Journal of Development Economics 99 (2012) 2745

Contents lists available at SciVerse ScienceDirect

Journal of Development Economics

journal homepage: www.elsevier.com/locate/devec

Evidence on the impact of minimum wage laws in an informal sector: Domestic

workers in South Africa
Taryn Dinkelman a,, Vimal Ranchhod b
a
b

Dartmouth College, United States

University of Cape Town, South Africa

a r t i c l e

i n f o

Article history:
Received 27 July 2010
Received in revised form 6 December 2011
Accepted 27 December 2011
Available online 4 January 2012
JEL classication:
J08
J48
O15

a b s t r a c t
What happens when a previously uncovered labor market is regulated? We exploit the introduction of a minimum wage in South Africa and variation in the intensity of this law to identify increases in wages for domestic workers and no statistically signicant effects on employment on the intensive or extensive margins.
These large, partial responses to the law are somewhat surprising, given the lack of monitoring and enforcement in this informal sector. We interpret these changes as evidence that strong external sanctions are not
necessary for new labor legislation to have a signicant impact on informal sectors of developing countries,
at least in the short-run.
2012 Elsevier B.V. All rights reserved.

Keywords:
Minimum wage
Informal sector
Domestic workers
Africa

1. Introduction
What happens to wages and employment in the informal sector after
the introduction of a minimum wage in that sector? The vast minimum
wage literature in economics has remarkably little to say on this question, since the informal sector has more often represented the uncovered sector in this research and has been used to help distinguish
between models of the labor market. For example, if the wage and employment effects of a change in the formal sector minimum are mirrored
by opposite-signed responses in the informal sector, this is consistent
with a segmented two-sector labor market model.1 However, if informal
sector wages increase and employment decreases when a formal sector
minimum is adjusted upwards, this is consistent with spillovers in an

Without implicating them in any way, this paper has beneted from discussions
with John Bound, Charlie Brown, John DiNardo, David Lam, James Levinsohn, Justin
McCrary, Jeffrey Smith and Gary Solon and from the helpful comments of Tom Hertz
and several referees. We are very grateful to Nicola Branson for sharing her data on
LFS sample weights with us.
Corresponding author.
E-mail addresses: Taryn.L.Dinkelman@dartmouth.edu (T. Dinkelman),
vimal.ranchhod@gmail.com (V. Ranchhod).
1
See Brown (1988) for a discussion of two sector models in minimum wage studies.
Brown (1999), Card and Krueger (1997) and Neumark and Wascher (1995, 2007) provide comprehensive reviews of the large minimum wage literature.
0304-3878/$ see front matter 2012 Elsevier B.V. All rights reserved.
doi:10.1016/j.jdeveco.2011.12.006

integrated labor market, or possibly with lighthouse effects.2 In this

paper, we extend the minimum wage literature and investigate whether
a minimum wage oor can have any effect when directly applied to the
informal sector, where the institutional environment for monitoring and
enforcement of penalties is weak. We use the introduction of the rst
minimum wage in South Africa's domestic worker sector in November
2002 to analyze what can happen in the short-run.
Whether a minimum wage can have any impact in the informal
sector is broadly relevant for understanding more about the process
of labor market formalization in developing countries. As economies
develop, labor relationships shift from rural to urban areas and take
place in larger and larger rms; employers and employees start to
pay taxes and workers gain legal protections, often including a guaranteed minimum rate of pay. Such protections, when enforced, may
signicantly improve working conditions and reduce poverty among
the least skilled workers (Lustig and McLeod (1997) summarize results from several studies), or may hinder the operation of markets
2
Spillover effects into the informal sector or other formal uncovered sectors have
been documented by many researchers, particularly in Latin American countries. See
Lemos (2009) for Brazil, Gindling and Terrell (2004) for Costa Rica, Maloney and
Menendez (2004) for a variety of Latin American countries, Bell (1997) for Mexico
and Colombia. See also Cortes (2004) for the USA. Brown (1988) notes that in the
US, a large fraction of workers earn the minimum wage, even though they are
employed by establishments not subject to the minimum wage law.

28

T. Dinkelman, V. Ranchhod / Journal of Development Economics 99 (2012) 2745

and negatively affect productivity (as in Besley and Burgess, 2004) and
employment.3 For many countries contemplating such regulation, resources for monitoring and enforcement are limited. And, since most
economic models of employer behavior require some penalty and a
non-zero probability of being audited to predict any effects of minimum
wage regulations (see Ashenfelter and Smith (1979) for the canonical
model of minimum wage compliance), a rst order question for many
developing countries is what effects can be expected from new labor
legislation in the face of limited monitoring and enforcement? 4
In this paper, we shed light on the effects of new labor regulation in a
context with little active enforcement and no clear penalties. We present
an empirical example in which domestic work employers face an extremely high minimum wage (set at the 70th percentile of the pre-law
wage distribution) with no effective penalties and a vanishingly small
probability of being audited, and show that employers still choose to respond to the law. We document immediate, large and partial adjustment
of wages upwards in the wake of the law and nd no statistically significant effects on the intensive or extensive margins of work, at least in the
short-run, sixteen months after the law is enacted.5 We also see evidence
of dramatic increases in the fraction of domestic workers who have a formal contract of employment, unemployment insurance coverage and
employer-provided pension contributions after the law. This South African case indicates that the effects of labor legislation may not rest on the
type of enforcement that exists in already formalized markets; rather,
the introduction of the law itself may serve as a focal point for shifting
markets in the direction of becoming more formal. 6
The domestic work sector is important in its own right, and is
under-studied given its prevalence over space and time. Historically,
this sector has been important in developed countries. Rubinow
(1906) uses census data to show that over 1.2 million women were
employed in domestic work in the US in 1900. World-wide, the market for domestic workers currently employs many millions of
women: foreign workers employed in private households make up
around 10% of the labor force in a number of Middle Eastern countries
(Kremer and Watt, 2006); foreign domestic workers constitute 6% of
the workforce in Hong Kong (Cortes and Pan, 2009); while ILO data
for OECD countries record an average of 100,000 female domestic
workers per country. The UK and Germany are at the high end of
this range with 400,000 and 460,000 female domestic workers respectively. In South Africa, about one in three working women are
employed in domestic work.
In most of these labor markets, the domestic work sector ts the ILO's
denition of informal: the majority of rms are one-employee enterprises (households) in which labor relations are predominantly uncontracted and workers do not enjoy minimum wage protections or other
non-wage benets like unemployment insurance or pensions. The

3
In that paper, Besley and Burgess (2004) show how active and costly pro-worker
regulation in the formal manufacturing sector in India led to increased informality, reduced investment and lower labor productivity.
4
Despite the greater availability of resources for enforcement in developed countries, noncompliance with minimum wage laws is widespread. Ashenfelter and Smith (1979) note that
compliance with the US Federal minimum wage was only 65% in 1973; Cortes (2004) reports
that in 1997, as many as 40% of US workers who qualied were paid less than the minimum
wage; and non-compliance rates in excess of 50% have been reported for Mexico, Morocco
(Squire and Suthiwart-Narueput, 1997) and other developing countries. See Neumark and
Wascher (2007) for a comprehensive review of the literature on minimum wages from developing countries. All of the theoretical literature on compliance with a minimum wage
hinges on employers choosing an optimal level of compliance in the face of penalties and enforcement. See Grenier (1982), Chang and Ehrlich (1985), Bloom and Grenier (1986), Chang
(1992), Lott and Roberts (1995) and Weil (2005).
5
One benet of focussing on short-run effects is that there is little time for workers
to sort across areas and relocate e.g. from rural to urban areas in search of higher wage
jobs. In our data, roughly the same fraction of domestic workers report starting their
current job in the past year, both before and after the law (12% in September 2001
and 14% in September 2003).
6
The rst minimum wage law introduced in the US in 1912 (in Massachusetts, for
women) had some similar features to our setting. One penalty involved newspapers
naming and shaming non-compliant rms by publishing their names (Thies, 1991).

small scale of employers makes this sector costly for unions to organize.7
Work often extends beyond simple housekeeping services and can require a great deal of trust, particularly when child-care is involved or
when the employer is absent during working hours. Additionally, and
particularly in Asia, Latin America and the USA, domestic workers are
often foreign migrant workers with tenuous legal status. This increases
their vulnerability in the labor market.8 For example, in the USA, domestic workers are largely undocumented and are not yet guaranteed all of
the protections of the National Labor Standards Act.9 We believe our
analysis of the South African case sheds light on the short-run effects of
introducing minimum wage legislation to these informal and uncovered
work relationships that are found throughout the world.
The empirical exercise in this paper is straightforward. We evaluate
the effects of South Africa's 2002 minimum wage law for domestic
workers by exploiting time-series variation in the application of the
law and pre-existing cross-sectional variation related to the intensity of
the law to identify wage and employment effects. We use labor force survey (LFS) data from 2001 to 2004 to capture worker-reported wages,
hours of work and employment every six months. Large shifts in the
wage and earnings distributions of domestic workers are evident in our
non-parametric kernel densities. Exploiting only the before-after variation, domestic worker wages increase by about 20% in the 16 months
after the law. Although this sharp jump in wages in a relatively short period of time is strongly suggestive of the new law having had an impact
in this sector, we are naturally concerned that contemporaneous shocks
to the economy, or differential economic trends might show up as similarly large wage increases. We complement the before-after analysis
with a difference-in-differences strategy that adopts the methods in
Lee (1999) to statistically examine the effects of the law.10 Specically,
we compare the change in wages and hours of work of domestic workers
in places where the median wage was far below the wage oor in the
pre-period (high wage gap areas), to places where the median wage
was closer to the minimum (low wage gap areas), thereby combining
cross-sectional and time-series variation in the application of the law.
We nd that wages increase by a statistically signicant 1315% in the
wake of the law. In contrast, we nd no statistically signicant reduction
in hours of work nor any signicant change in the probability of a lowskilled female worker being employed as a domestic worker in the preversus post-period, in high wage gap compared to lower wage gap areas.
Although the minimum wage law as it was enacted exposed all
urban areas at the same time, we make use of the fact that the law
is more demanding of employers in urban areas with lower pre-law
wages. 11 And, as in any difference-in-differences research design
that exploits a change in policy at one point in time, a key identication assumption is that both the exposed and unexposed groups are
7
The ILO denes informal employment as all remunerative work (i.e. both selfemployment and wage employment), that is not registered, regulated or protected
by existing legal or regulatory frameworks, as well as non-remunerative work undertaken in an income-producing enterprise. Informal workers do not have secure employment contracts, worker's benets, social protection or workers' representation
(http://www.ilo.org/public/libdoc/ILO-Thesaurus/english/tr1746.htm).
8
This idea of heightened vulnerability is not new. In 1906, Rubinow (1906) writes
about the servant girl's problem and describes American preferences for hiring foreign women for domestic work as being related to the greater ease of managing
them, which translates into longer hours, perhaps lowers wages, more work and,
in general, conditions of service more favorable to the employer.
9
The state of New York recently became the rst state in the USA to sign into law a
Domestic Workers Bill of Rights. Among other rights, the new law ensures that domestic workers have notice of termination, receive paid sick days and holidays, and other
basic labor protections that are standard in the Fair Labor Standards Act. See the editorial The Rights of Domestic Workers, The New York Times June 15, 2009. http://www.
nytimes.com/2009/06/15/opinion/15mon3.html Also the article Senate Passes Historic Bill To Protect Domestic Workers at http://www.nysenate.gov/press-release/
senate-passes-historic-bill-protect-domestic-workers posted on June 2, 2010.
10
In that paper, Lee (1999) uses regional differences in the relative level of the US
federal minimum wage to identify the effects of the minimum wage law on wage inequality in the 1980s separately from the effects of national trends in wages.
11
Although there is some spatial variation in the level of the minimum, we explain in
the next section why we do not use this variation.

T. Dinkelman, V. Ranchhod / Journal of Development Economics 99 (2012) 2745

on the same trend in the absence of the policy change. This is particularly important for economic variables like wages and hours, that are likely
to be able to move quickly in response to changes in general economic
conditions. An important defense of our identication strategy therefore
relies on providing some evidence that areas with larger pre-law wage
gaps were not simply experiencing faster trend growth in wages and
were unlikely to be exposed to contemporaneous positive economic
shocks. We address these concerns in two ways. First, we show that
areas with the smallest pre-law wage gaps appeared to experience the
fastest GDP growth over the period, suggesting that if anything, labor
demand trends are stronger in these less intensely treated provinces.
Second, we perform a placebo experiment using a set of similar workers
who are unlikely to compete over jobs with domestic workers, but
whose job conditions likely reect general economic conditions: lowskilled male manufacturing workers in urban areas. We nd that male
manufacturing wages do not grow faster in periods after the law in the
places where the domestic worker minimum wage was more binding.
This gives us more condence that the difference-in-differences results
we estimate for domestic workers are not being driven by differential
wage trends between high and low wage gap areas. This sample of
male manufacturing workers is also helpful for a second reason, as
they also allow us to rule out the possibility of strong mean reversion
in wages as a reason for our positive domestic worker wage results.
Quite apart from the wage and employment effects of the law, we
document that the introduction of the law had a substantive impact
on more general conditions of work for domestic workers. We examine
the probability that a domestic worker is protected by a formal job contract with their employer, or has an employer providing unemployment
insurance or pension benets after the law. The probability of an employee having a formal contract more than doubles in the sixteen
months after the law, regardless of the intensity of the minimum
wage oor in their area of work; the fraction of workers enjoying pension benets increases by about 7 percentage points, and the fraction
of workers having UIF contributions made for them increases by 18 to
20 percentage points. These improvements in the conditions of work
for domestic workers are substantial, immediate, and importantly, occur
across the distribution of wages (i.e. not just in areas where workers
were originally earning below the minimum). These results suggest the
beginning of the formalization of this industry, with potential farreaching consequences for the nature of domestic work in South Africa.
Given the weak institutional environment for enforcement of the law,
which we describe, it is somewhat surprising that we see such large wage
responses and no employment effects in this informal sector. Isolating the
exact reasons for the employer response is not easy; however, we propose two pieces of evidence that suggest that employers were voluntarily
and only partially responding to the law. First, we develop a test for partial
compliance, which shows that some workers get increases bringing their
wages closer to, but not nearly up to, compliant levels. This partial compliance likely contributes to the lack of employment effects of the law.12 Second, we show that the wage response of employers is not signicantly
different across places with different audit probabilities, where we use
the presence of a local Labor Center (LC) as a proxy measure of this probability. Partial compliance that does not appear responsive to the likelihood of audit is consistent with the idea that employers may not have
been primarily motivated by the threat of external sanctions.
Two related, unpublished papers have examined the effect of the
minimum wage law for domestic workers in South Africa: Hertz
(2005) and Yamada (2008). Our study differs from these studies in several ways. First, we focus on urban workers only, because we are concerned that identication of the impact of the new minimum wage
law for rural domestic workers is likely confounded by the concomitant
introduction of a minimum wage for agricultural workers, a plausible
12
This nding relates to a theoretical point made by Basu et al. (2007). The authors
develop a model in which governments accept some non-compliance with minimum
wage legislation to achieve distributional goals.

29

alternative sector of work for low-skilled workers in rural areas.13 Second, we use an updated and consistent set of survey weights for the LFS
data (Branson, 2009) which were not available for these studies.14
Third, we use a higher level of aggregation (the province) to dene
areas in which the new law was more or less binding, based on prelaw characteristics. This choice presents other challenges for inference
that we deal with using an appropriate two-step estimator, described
in detail in our empirical analysis section. 15Fourth, we present evidence
from a placebo test using outcomes for male manufacturing worker that
bolster our causal claims for the effects of the law on domestic worker
wages; this sample also allows us to rule out the possibility that mean
reversion could account for the wage effects. Fifth, our results differ
from Hertz (2005) and Yamada (2008). Although we also nd positive
wage effects of the minimum wage law, we nd little evidence for a statistically signicant negative employment effect of the law on either the
intensive or extensive margins of work. Finally, we also try to understand more about the motivations for response to the law. We show
that some employers appear to make partial adjustments of wages towards the minimum wage, but not quite up to the legal wage oor.
And, we show that the wage effects of the law do not seem to be driven
by employers responding to a potential threat of audit, as proxied by the
presence of labor centers in the district. Our interpretation of this evidence is that the law sparked the beginning of a formalization of this
market (as also evidenced by the improvement in contract coverage,
UIF and pension benets) without strong external sanctions.
The paper begins with a description of the domestic worker industry
in South Africa before the law and describes the characteristics of the
new law introduced in 2002. After describing the data and presenting
summary statistics of our sample, we turn to documenting the wage,
earnings, employment and hours of work effects of the law using a combination of kernel densities for wages and earnings and difference-indifferences regressions. We discuss our test for and present evidence
of partial compliance with the law. We show that wage increases are
not larger in places with a higher probability of being audited, and
that the law increased the probability of a domestic worker having a
formal employment contract, UIF and pension benets regardless of
how large the initial wage gap in the province of residence was. We conclude with a discussion and interpretation of the results.

2. The domestic worker industry in South Africa

The domestic worker industry in South Africa employs 18% of all
women, and 80% of all domestic workers are female. Poorly educated African and colored women make up the vast majority of these domestic
workers.16 In each year of our study, about 35% of urban African and colored female workers were in the domestic work sector, and about 60% of
all domestic workers were employed in urban areas. Unlike many Latin
American and Middle Eastern settings, and more like countries in the
rest of Africa and parts of India, the majority of domestic workers in
South Africa are not foreigners. Table 1 presents means and standard deviations of female domestic worker demographics for the period before
the law (September 2001, March 2002 and September 2002) and the
period after the law (March 2003, September 2003 and March 2004).
13
Agricultural workers received protection under a minimum wage law, also for the
rst time, six months after the domestic worker minimum wage oor was imposed.
14
Details of this choice are discussed in the data section.
15
Hertz's (2005) working paper uses some of the same LFS cross-sectional data to estimate the impact of the law on wages and employment using a difference-indifferences approach that relies on a much smaller unit of analysis (the magisterial district).He relates the intensity of the minimum wage law to the fraction of workers in a
magisterial district who initially earned below the minimum (following Card and
Krueger, 1997). We choose not to use magisterial districts as the unit of analysis, since
many districts contain only a few (under 10) individuals in each wave who are
employed as domestic workers.
16
Following much of the economic literature on South Africa, we use apartheid-era racial
classications: African for Black South African, and colored for individuals of mixed race.

30

T. Dinkelman, V. Ranchhod / Journal of Development Economics 99 (2012) 2745

Table 1
Sample summary statistics.
N

Pre-law mean (s.d.)

Post-law mean (s.d.)

PostPre difference (s.e.)

P value of the difference

(1)

(2)

(3)

(4)

(5)

A: Conditional on being in the labor force

Employed at all

52,739
52,739

0.64
(0.48)
0.34
(0.47)

0.01
(0.01)
0.01
(0.01)

0.26

Employed as a domestic worker

0.63
(0.48)
0.35
(0.48)

546.26
(386.73)
3.67
(2.91)
39.44
(14.75)
0.29
(0.45)
0.79
(0.41)
0.10
(0.30)
0.03
(0.18)
0.02
(0.15)

658.39
(451.04)
4.37
(3.59)
38.72
(14.10)
0.42
(0.49)
0.79
(0.40)
0.27
(0.44)
0.10
(0.30)
0.21
(0.41)

112.14
(26.08)
0.70
(0.17)
0.72
(0.82)
0.13
(0.02)
0.01
(0.02)
0.18
(0.02)
0.07
(0.01)
0.19
(0.02)

40.47
(9.29)
6.96
(3.41)
0.89
(0.31)

0.14
(0.24)
0.35
(0.08)
0.02
(0.02)

B: Conditional on being employed as a domestic worker

Nominal monthly earnings (ZAR)
6160
Nominal hourly wage (ZAR)

6154

Hours of work per week

6876

Fraction paid > = minimum

6155

Full-time worker

6876

Fraction with a job contract

6784

Fraction with a pension

6867

Fraction with UI coverage

6867

C: Characteristics of women employed as domestic workers

Age
6876
40.33
(9.37)
Education (years)
6876
6.62
(3.44)
African
6876
0.90
(0.29)

0.13

0.00
0.00
0.41
0.00
0.88
0.00
0.00
0.00

0.59
0.00
0.42

Data are from South African Labor Force Surveys (LFS 20012004). Sample includes African and colored females aged 2059 inclusive, who have no more than a completed high
school education and who live in urban areas. Panel A includes workers and unemployed women looking for work; Panels B and C restrict to women employed as domestic workers
in any period. All statistics are weighted and the standard errors of differences and p-values are calculated taking these weights and province-level clustering into account. The prelaw period includes LFS waves in September 2001, March 2002 and September 2002; the post-law period includes LFS waves in March 2003, September 2003 and March 2004. A
full-time worker is someone who reports at least 27 hours of work per week; UIF is unemployment insurance.

All statistics are weighted, and the data sources are described in more
detail in the next section.17
The average age of these workers is around 40, the majority are
African women and they have between 6 and 7 years of education,
which is roughly completed primary school. This is 0.8 to 2 years
below the average education of women working in the most closely
related skill group: women in elementary occupations (e.g. newspaper vendors, ofce cleaners, hawkers, building caretakers, garbage
collectors etc) and the female self-employed. 79% of domestic
workers report working full-time, dened as 28 hours or more per
week, making the majority subject to the full-time minimum wage.
Domestic workers are typically poorly remunerated. Mean wages
are lower in this occupational category than in any other: the ratio
of the mean domestic worker wage to the mean wage for other
low-skilled African and colored elementary workers (self-employed
women) was 0.49 (0.64) in September 2001. Prior to November
2002, there was no minimum wage in the domestic worker sector
and no formal mechanism existed for domestic workers to negotiate
wages. Wages were typically set unilaterally by the employer household or in consultation with other local employers (see Cock (1989)
for a qualitative description of this process). Although some aspects
of the 1997 South African Basic Conditions of Employment Act governed overtime provisions, leave considerations, minimum notice periods, fair dismissal procedures and severance pay for all workers (South
17
Throughout, we use the following denition of domestic workers: currently
employed African or colored females aged 20 to 59 inclusive, who live in urban areas
and have their occupation coded as working as a domestic worker in a private household for the week prior to the survey. We exclude a handful of these workers who also
report having their own business on the side and those who have more than a high
school level of education (12 years).

African Department of Labor, 1997), these were rarely adhered to among

domestic worker employers (Louw and Van der Berg, 2004). For example,
only 10% of domestic workers had a formal contract of employment in
2001, compared with 55% of elementary occupation workers.
In setting the rst national minimum wage for domestic workers, the
Department of Labor (DoL) took into account the recommendations of a
government-appointed Employment Conditions Commission. This
group of government representatives and academics dened the scope
of the Domestic Worker Sector and concluded that any wage oor should
improve the livelihoods of those worst off and retain jobs. Their recommendation for the actual minimum wage level was higher than that
initially proposed by the government, and was the one eventually
adopted (Budlender et al, 2002).
Under the new law, which became effective on 1 November 2002, domestic workers and gardeners working in private homes had the right to
a minimum wage and to 8% annual wage increases.18 The urban full-time
hourly minimum wage was set at ZAR4.10 (USD 0.410) in November
2002; the part-time wage was ZAR4.51 (USD 0.451) where part-time
work is dened as fewer than 28 hours per week.19 Since about 80% of domestic workers in urban areas work full-time, we focus on the urban fulltime minimum as the relevant wage oor throughout the paper.20 In
18
Garden workers, most of whom are men, are also covered as domestic workers under
this law. However, they make up a minority of domestic workers and we omit them from
our analysis.
19
The average Rand/USD exchange rate from June 2002 to January 2003 of
ZAR10 = USD1.
20
The law species different wages for two types of urban areas, Area A and Area B
localities. These areas generally differ in size, and since Area A localities are the largest
urban parts of the country, we use the full-time wage set for these areas as the urban
minimum wage. We cannot distinguish between A and B areas in our data to create a
ner measure of treatment.

T. Dinkelman, V. Ranchhod / Journal of Development Economics 99 (2012) 2745

addition, the new law enabled employers to deduct up to 10% of the total
salary for rental value of any accommodation provided. A separate and related piece of legislation, introduced shortly afterwards on April 1, 2003,
additionally required employers to register domestic workers with the
DoL in order to pay unemployment insurance (UI).
It is important to note that the initial change introduced in this industry in November 2002 was an order of magnitude larger than typical changes in the value of the minimum that economists typically
study: the wage oor was set at 1.5 times the median monthly earnings of domestic workers in 2002. Full compliance in this context
would have entailed massive wage increases for a majority of workers
and potentially large negative employment effects for most employees either on the extensive or intensive margin. Given existing
high levels of unemployment in South Africa, this would be one possible reason for why the government did not commit substantial resources towards enforcement of the law.
In fact, in the rst ten months after the law, both the audit probability
and penalty for rst time violators were very small.21 As far as we have
been able to establish, no inspections were carried out until August
2003, ten months after the law. At this time, 1600 households in ve provinces were earmarked for inspection and only 25% of them were found to
be in compliance with the law.22 Although we have not been able to obtain ofcial aggregate statistics on household inspections during these
years, interviews with several labor centers from around the country indicate at most a couple of hundred household inspections per year.23
There were also no documented rules about penalties or back-pay
for non-compliers at the time of implementation. Press releases from
the DoL in February 2003 indicated that Non-compliance with the UI
law will result in penalties of up to ZAR5000 (USD500) per household
or ve years imprisonment.24 However, we have not been able to
nd ofcial documentation of this or other penalties, nor has our search
of newspaper archives revealed any reports of nes or prison sentences
being imposed on non-compliant employers.25 What we do know is
that non-compliers might have expected three progressively more
threatening warnings (telephonic, written, court order) before appearing at a court of law. At this time, the right of appeal would have been
well exercised, as it is unclear how evidence for non-compliance could
be substantiated in this predominantly cash payment industry.
All evidence from the DoL website and various legal documents
and reports related to the law suggest that the general monitoring
and enforcement regime in the domestic worker industry was weak
and presented employers with an almost zero expected cost of noncompliance. Despite this lack of compliance incentive, the timing of
21
During this time, employers might have expected a vanishingly small audit probability for two reasons. First, the chances of random inspection are small since inspections are labor intensive and each household yields only a single worker inspection.
Bhorat et al (2010) provide evidence of only one labor inspector per one million
workers for the entire country. If each domestic worker works in only one household,
this yields over one million employers that are subject to the minimum wage law in
this industry alone. Second, logistical difculties in gaining access to employer premises make any inspection costly. A non-compliant employer can legally refuse to allow an inspector into their private residence, or simply not be present at the time of
the inspection. A court order from a Labor Court is then required to enter the residence.
There are also physical barriers to entry: inspectors have reported difculties with impenetrable gates and the presence of dogs (Ofcial release by the DoL, 27 th August
2003. Available at www. labour. gov. za).
22
See media release at www.labour.gov.za/media/statement.jsp?statementdisplayid=
9685
23
We conducted a snap survey of the 104 Labor Centers in the country and found
that in several of them, between 100 and 250 household inspections were carried
out each year after 2004.
24
http://www.info.gov.za/speeches/2003/03050809461001.htm
25
The Basic Conditions of Employment Act (1997) states that underpayment violations for any worker are penalized in the following manner: rst offense25% of the
gap plus interest; second offense within 3 years50% of the gap plus interest; third offense within 3 years or second offense within 2 years75% of the gap plus interest;
fourth offense within 3 years100% of the gap plus interest; fth offense within
3 years200% of the gap plus interest. However, as noted in Louw and van der Berg
(2004), the BCEA conditions were seldom adhered to in informal sectors of work.

31

the law coincides with substantial rightward shifts in the wage and
earnings distributions of domestic workers. We next describe the
data which we use to document these shifts.
3. Data description and empirical methods
3.1. LFS surveys
We use six cross sections of data from the nationally representative South African Labor Force Surveys (LFS): September 2001,
March 2002, September 2002, March 2003, September 2003 and
March 2004. These LFS surveys are biannual rotating panel surveys,
conducted in February/March and September each year and include
detailed data on the work and unemployment experiences of 60,000
to 70,000 working-age individuals living in 30,000 households. The
six waves we use span the period just before and just after the minimum wage law becomes effective in November 2002. The survey instrument is similar to the US Current Population Survey, although the
rotation pattern differs. In each wave, 20% of households interviewed
in the previous wave are rotated out of the survey entirely. 26 These
LFS data are high frequency and can be used to examine differences
over a six month window. They help us to estimate the immediate
impacts of the law while controlling for observable characteristics of
domestic workers (i.e. age, race and years of education) and to see
whether pre- versus post comparisons are sensitive to these controls.
We also model the probability of being employed as a domestic worker in each wave, using the sample of employed and unemployed
women with similar age and education proles to domestic workers.
We use these LFS data as repeated cross sections and exploit the
cross-sectional variation in intensity of the law at the province level
(9 provinces in total) in combination with the time-series variation
in the application of the law to identify the effects of the law. 27
3.2. Sample weights
One further aspect of the data that is worth noting concerns the appropriateness of the sample weights available in the LFS. Survey weights
for LFS 2001 and 2002 are benchmarked to the 1996 Census; while survey weights for LFS 2003 are benchmarked to the later 2001 Census.
With different benchmark years, this series of weights is potentially inconsistent over time. Hertz (2005) provides a detailed discussion of how
these inconsistencies in the weights may affect any analysis of the minimum wage law. He points out that the 2001 Census under-counted the
fraction of women of working age and that extrapolations between 1996
and 2001 overestimated growth in the adult population by underestimating the effect of the HIV epidemic on adult totals. He notes that
no consistent ofcial series of sampling weights is available. This
poses a serious problem, as both the changes in scale and the changes
in the age, gender, province, and race group distributions result in artifactual changes in the measured employment of domestic workers
26
Although there are three earlier waves of data going back to 2000, the baseline
sample was drawn anew for the September 2001 round which is why we begin our
analysis with data from this round. And, although the LFS survey has continued biannually since March 2004, we cannot use additional waves of post-data in our analysis
since the survey stopped reporting whether the individual resides in an urban or a rural area, thus making it impossible for us to condition our sample on urban domestic
workers (Statistics South Africa, 20002003).
27
There is a panel data component of the LFS survey, but we have some concerns
about the representativeness and quality of the panel data set of workers. First, since
the design of the panel includes a 20% out-rotation of dwellings in each six month period, in order to appear in the panel a worker needs to be living continuously in the
same place for six waves and needs to escape the out-rotation group. We expect that
workers who are more likely to retain employment are also more likely to appear in
the panel, making the estimation of the effects of the law on such a selected sample difcult. Second, we learned from Statistics South Africa (the organization that collects
the LFS data) that some panel matches could not be made, because some fraction of
questionnaires from the pre-period were lost in a ood. We have no way to know what
impact this would have on the representativeness of the panel.

32

T. Dinkelman, V. Ranchhod / Journal of Development Economics 99 (2012) 2745

that are too large to be ignored. Indeed, the original sample weights
that he refers to have been shown to produce inconsistent aggregate statistics over time. More formally, Branson (2009) explains that The
StatsSA weights presented in the data are problematic for analyses
over time the auxiliary data used as a benchmark in the poststratication adjustment are unreliable and inconsistent over time and
hence result in temporal inconsistencies even at the aggregate level.
For these reasons, we do not use the weights provided with the
LFS data. Rather, our results use new (individual-level) survey
weights constructed in Branson (2009) using entropy estimation. 28
As shown in that paper, the weights produce consistent demographic and geographic trends.
3.3. Sample selection and key variables
Our main analysis sample includes all urban African and colored
woman aged 20 to 59, who report domestic work in the week before
the survey, who do not also own their own business and who have no
more than high school education. For the employment analysis, we
use an expanded sample of all African or colored woman aged 20 to 59
who live in urban areas, who have no more than a high school level of
education and who are employed or looking for work. For the placebo
test, we make use of a sample of male urban African and colored
workers, aged 20 to 59, who have no more than a high school education
and who report working in the manufacturing sector.
In the LFS, all workers are asked about earnings, pay frequency and
usual weekly hours of work. Most workers report earnings and a corresponding pay frequency. The vast majority 89% of domestic workers
report a monthly pay frequency. We convert all earnings to monthly
amounts using the pay frequency information. About 89% of domestic
workers in any one wave do not report earnings or only report earnings
in brackets; we exclude these individuals from our analysis.29 To capture
hours of work, we use the response to the question How many hours do
you usually work in a week? We construct hourly wage measures by dividing monthly earnings by (Usual hours worked per week) (Average
weeks in a month). Just over 5% of workers report working more than
70 hours a week and we exclude these outliers from our analysis.
3.4. Empirical strategy
We describe the impact of the minimum wage law in two ways.
First, we estimate non-parametric kernel densities of domestic worker wages and earnings and test for differences in the domestic worker
distributions over time using KolmogorovSmirnov tests of the
equality of distributions. Then, we statistically test for whether outcomes changed in the period after the law compared to before the
law, and for whether these changes are larger in areas where the minimum wage initially had more bite. We specify the following
difference-in-differences regression model:
yijt 0 1 POST t 2 WGj 3 POST t  WGj X ijt ijt

where yijt is one of several main outcome variables for individual i living in province j in period t: log hourly wages, weekly hours of work,
the possession of a formal job contract for the set of domestic
workers, whether the individual gets pension or UI benets, and
whether the individual is employed as a domestic worker or not
among the set of demographically similar women.
28
These methods are described in Branson (2009). Entropy estimation essentially
creates a new set of weights that are as close to the original weights as possible (to preserve the main features of the sample design) but that are adjusted to account for errors arising from time inconsistent benchmarks, from inconsistencies between
household and person weights, and for errors in the trimming of weights in earlier survey years. We are grateful to Branson for making these weights available to us.
29
The fraction of domestic workers with no earnings information is not signicantly
different across waves.

Eq. (1) is estimated with and without controls for worker characteristics (Xijt), including age, years of education and an indicator variable for whether the person is African. We group data from the
September 2001, March 2002 and September 2002 surveys into a
Pre period and the remaining three waves into the POST period.
Because the law was formally announced on August 30, 2002, but employers were only expected to be compliant from 1 November 2002,
there is some ambiguity about whether the September 2002 wave belongs in the pre- or post-period. There was substantial publicity in the
months prior to the law becoming effective, making it plausible that
some employers started to react even before the November 1st deadline. For this reason, we also present the main results from regressions that omit the September 2002 cusp wave, for which POSTt is
not well-dened. These are our preferred estimates.
To construct a measure of the intensity of the minimum wage law,
we follow other examples in the minimum wage literature (notably
Lee, 1999) that use aggregate (rather than individual) data on
workers over time and construct a locally-specic wage gap using
data on wages from the PRE period. We dene the province-level
wage gap as:
h

i
WGj log minw log median wj

where min(w) is the urban full-time minimum wage in November 2002

and median(wj) is the median wage of all urban domestic workers in the
province before the law. That is, the intensity of the minimum wage law
going forward in time is measured according to domestic worker conditions in local labor markets in September 2001 and March 2002.
Provinces with very low median wages prior to the law therefore
have a large positive value for WGj. To aid interpretation of the coefcients, we implement this difference-in-difference regression with demeaned versions of the WGj and WGj*POSTt variables and all continuous
control variables.
Under the assumption that the wage gap measure (and therefore
the difference-in-differences term) is orthogonal to the error in (1),
the parameter 1 tells us how outcomes changed on average across
all areas after the law while 2 tells us the average difference in outcomes for domestic workers in areas with larger than average wage
gaps, across the entire period. Of course, any general economic trends
that affect outcomes in the post-period would also be part of the 1
parameter. 3 is the difference-in-differences parameter: it tells us
how much more outcomes changed after the law, in areas where
the minimum wage was more binding.
There are two main concerns with using the median wage gap
measure as our measure of the treatment intensity of the law. First,
there is the obvious worry that high and low wage gap areas may
have been trending differently in the POST period, in a way that confounds the effects of the law. Second, there is the potential for mean
reversion in wages to account for the wage results.
To deal with the rst concern, we do two things. First, we show
that provinces with larger wage gaps (i.e. more intensely treated
areas) do not appear to be growing faster over time, relative to
small wage gap provinces, making differential trends in labor demand
an unlikely explanation for our wage results. Second, we show for a
specic subset of workers that there is no evidence of positive contemporaneous shocks to large wage gap provinces that could account
for our results. We implement a placebo test by estimating the same
regression model in (1) for a set of workers for whom the minimum
wage law is irrelevant and who do not directly compete with domestic workers for jobs: male, low-skilled manufacturing workers in
urban areas. The minimum wage law for domestic workers does not
apply to them and, moreover, is set at a level far below their median
wage. Therefore, if we see manufacturing worker wages increasing
more in areas where domestic worker wages were substantially

T. Dinkelman, V. Ranchhod / Journal of Development Economics 99 (2012) 2745

below the minimum, we would be concerned that all wages in high

wage gap provinces were increasing over time. This is not the case.
To address the second concern about mean reversion in wages, we
look for evidence of this in the manufacturing worker sample. We
alter the denition of WGj to be:
h


i
WG2j logminwlog median wj manuf acturing

where we use the domestic worker minimum wage as an arbitrary

benchmark, and the median manufacturing worker wages in the period before the law to construct the differenced variable. We implement
the difference-in-differences regression using this alternate measure
of treatment. If there is strong mean reversion in wages, we should
see manufacturing worker wages increase in places with higher values
of WG2j, after the law.
A third concern relates to the main unit of analysis that generates
our identifying variation. Note that the measure of the intensity and application of the law, WGj, is captured at the provincial level. This means
that in estimating (1), we exploit variation at the group (province) level
to identify the effects of the law. The fact that there are only nine groups
(provinces) presents a challenge for inference. Consider the following
error components model for the error term ijt in (1):
ijt vj
ijt

Standard OLS without adjustment of the standard errors treats each

individual as if they contribute equally and independently to the variation in outcomes. Since = v +
and vj is the same for each individual in the same province, this approach overstates the amount of
variation in our data. The situation is not quite as dire as only having
nine observations, since we also have individual-level control variables
that differ within provinces and which account for a large part of the
variation in outcomes. However, it is clear that some adjustment for
the grouped nature of the error term is called for.
We take two approaches. First, we follow the common recommendation in the literature to estimate Eicker-White clustered standard errors
at the level of the province.30 However, the standard asymptotic arguments for the consistency of clustered standard errors may not apply
with the small number of groups in our context, even given the additional variation found in the demographic controls. We still run the
risk of underestimating standard errors (and over-rejecting the null)
using this approach.31
As a second approach, we follow Donald and Lang's (2007) suggestion
and implement their two-step estimator which, under some conditions,
produces standard errors that appropriately take into account the
group-specic term in Eq. (4). To implement this, we rst regress outcomes on all individual level variables, the POSTt variable, a full set of
province dummy variables and a full set of POSTt*province interaction
dummy variables. Then, we take the estimated coefcients and use
them as the outcomes in a second stage regression. This second stage involves regressing the estimated coefcients on a POSTt indicator, the WGj
measure and the interaction POSTt *WGj. The resulting standard errors
from this second stage model are calculated with the group-component
of taken into account and, together with the second stage coefcients,
form t statistics that have the t distribution when the number of groups is
small.32 There is a nice intuition for this estimator: the rst stage
30
This approach to estimating appropriate standard errors in a difference-in-differences
specication is discussed in Bertrand et al. (2004) and in Donald and Lang (2007).
31
In Appendix 3 Tables 1 and 2 we also present block bootstrapped standard errors,
where the province is the block. Results do not differ substantially from the OLS results,
in terms of which variables are statistically signicant.
32
The conditions required for this result are that the number of individuals in each
province is large and that the underlying vj's are normally distributed. We must assume the latter condition, and having at least 600 observations in each group gives
us some condence in the former condition. For more details on this procedure, see
Donald and Lang (2007).

33

regression produces estimates of the group-level means in the postperiod (these are the coefcients on the province indicator variables
and the POSTt*province interactions) after taking into account variation
in the other individual controls. In the second stage, we estimate how
much of this variation in these group-level (estimated) means is predicted by variation in POSTt, WGj and POSTt *WGj.33
In addition to the graphical evidence from the kernel densities and
the difference-in-differences statistical tests, we implement two tests
to investigate the mechanisms through which the new minimum
wage law had an effect. We are particularly interested in understanding whether or not employers are complying with the law primarily
because they are concerned with penalties and external enforcement.
Our rst test shows that there is only partial compliance with the
law. That is, some employers are responding to the law, but their response is insufcient for compliance. In the absence of individuallevel panel data which we could use to test whether a worker experiences only partial wage adjustment to the law, we devise a related
test for partial compliance using repeated cross-sections. First, we
classify domestic workers into each of three bins, where the near
compliant bin consists of workers earning the minimum or up to
10% below the minimum, the compliant bin contains all workers
who earn above the minimum and the non-compliant bin contains
workers that are earning more than 10% below the minimum. We estimate an ordered probit model of this new variable on a constant, a
POSTt indicator and a set of demographic and geographic controls
(in a rst specication) and then also include controls for the wage
gap and the POSTt * WGj interaction (in a second specication). We
then predict the relevant marginal effects to estimate the change in
the probability of being in the near compliant bin and test for whether this change is positive. Appendix 1 sets out the details of how we
derive this test as the relevant one for partial compliance. The intuition behind the test is that if the probability of workers reporting
wages in the non-compliant bin falls by more than the probability
of workers reporting wages in the compliant bin rises, then some
workers are earning more than they were before (shifting to the
near-compliant region of the wage distribution), but not enough
more to make them compliant with the law. Such a result would be
consistent with an argument whereby employers are responding,
but not because they are law-abiding. Note that this interpretation
of the test relies on there being no large employment adjustment to
the law, a fact that our earlier estimates demonstrate.
Our second test shows that the size of the wage effect is not related to a proxy for the probability of being audited by the DoL. Although
we have described above how the probability of being monitored was
very small, it is still possible that domestic workers themselves may
have threatened to report non-compliant employers to the authorities. Indeed, Hertz (2005) reports an immediate increase in the number of domestic worker complaints made to the Commission for
Conciliation, Mediation and Arbitration (CCMA) after the law is introduced. The relevant labor authorities we consider here are housed in
local LCs. Under the assumption that it is easier to visit an LC in order
to lodge a complaint if there is an LC in your local labor market, we
can use the presence of an LC in the worker's local labor market as a
proxy for a higher probability of audit. We control for this proxy
and its interactions with POSTt, WGj and POSTt * WGj. We treat the
magisterial district in which the domestic worker resides as the
local labor marketa more disaggregated measure than the province.
Although we have not been able to obtain quantitative evidence on
the probability of being audited in relation to distance from an LC,

33
In the absence of demographic or other controls that vary within province, this
two-step estimator is equivalent to collapsing the data to the province-year level and
estimating a form of the difference-in-differences specication on these provincelevel means. However, we use the two-step estimator here to take advantage of the
fact that we have individual level controls that do matter for outcomes, while not overstating the amount of policy variation we have to identify the impact of the law.

T. Dinkelman, V. Ranchhod / Journal of Development Economics 99 (2012) 2745

.2

.4

.6

.8

34

-2

0
2
log hourly wage for urban domestic workers
Sept 01 - before
Sept 02 - before
Sept 03 - after

Mar 02 - before
Mar 03 - after
Mar 04 - after

Fig. 1. Distribution of log hourly wages for urban domestic workers. Kernel density plots of log hourly wages (bandwidth 0.02). Data are from South African Labor Force Surveys
(March 2001March 2004). The vertical line is at the level of the full-time minimum wage (monthly income) for urban domestic workers. Each wave of data contains between
996 and 1260 observations. KolmogorovSmirnov tests of equality of distributions reject at the 5% level for each pairwise comparison of waves in the before and after periods.

several LC's reported to us that the majority of their complaints from

this sector are received in person; moreover, LCs are legally and institutionally responsible for investigating complaints about working conditions and responsible for the enforcement of sectoral determinations
across all industries (Bhorat et al, 2010).34 If employers respond to the
law because they fear external enforcement, we should see wages rising
by more in places with an LC, compared to places without one.

4. Results
4.1. Before-after and difference-in-differences comparisons
The introduction of a minimum wage in the domestic worker industry appears to have had immediate and substantial effects on
earnings and wages of the average domestic worker, yet limited effects on hours of work. Table 1 shows our initial evidence from prepost comparisons of means. Panel B presents the means and standard
deviations of several outcome variables for domestic workers before
the law (column 2) and after the law (column 3) as well as the difference (standard error of the difference) in these means and the p-value
of the difference in the nal columns.
The rst point to note is the large jump in mean earnings and
wages from before to after the law: both monthly earnings and hourly
wages increase by about 20% and this difference is statistically significant. After the law, the average wage of domestic workers is higher
than the minimum, at ZAR4.37 per hour whereas before, the average
worker was being paid signicantly below the minimum at ZAR3.67.
The fraction of workers paid above the minimum increases from
under one-third prior to September 2002 to over 40% after the law.
Another striking result from the table is that the variance of wages
and earnings in the post period illustrates a signicant increase in dispersion relative to the pre-period. This is unusual, since wage compression is typically observed in response to increases in binding
minimum wage laws. This suggests that either some employers at
34
Several labor centers reported to us that all citizens are able to report complaints
directly to them, either by phone or in person, and that both methods are used. These
are also the units responsible for doing communications outreach with workers and
employers and for conducting mass inspections across several workplaces, in all
industries.

the lowest end are not responding to the law at all, or that employers
at the high end are also increasing wages even though they are already compliant with the law, or both.
Figs. 1 and 2 underscore the ndings in Table 1 and show the
movement of the entire wage and earnings distribution over time.
Each gure is a kernel density-smoothed plot of the earnings and
wage densities in September 2001 (14 months before the law),
March 2002 (8 months before the law) September 2002 (2 months
before the law), March 2003 (5 months after the law), September
2003 (10 months after the law) and March (2004) (16 months after
the law). The vertical line represents the urban full-time minimum.
Pre-law, there is no evidence in the graph that earnings are shifting,
despite annual ination rates of 6 to 7%. 35 In fact, in these preperiods, mean domestic worker wages are declining in nominal
terms. However, each gure shows that the entire (log) earnings
and wage distributions shift to the right in March 2003 with a pronounced mass in the lower tail moving towards the minimum wage
line. The shift is even more pronounced by September 2003 and
March 2004. We test for signicant differences between these distributions using KolmogorovSmirnov tests in all pairwise comparisons
of each wave and nd that each of the post-law distributions is significantly different from each of the pre-law distributions. In addition to
the prominent shift in wages and earnings after the law, it is clear
from these gures that a large fraction of domestic workers continue
to earn less than the minimum, and that although there is a modal
wage, there is no sharp minimum wage cliff that is characteristic of
US data for low wage workers (Dinardo, Fortin and Lemieux, 1996).36
Turning back to Table 1, there is little evidence of a large disemployment effect after the introduction of the law: hours of work fall
by 0.72 hours, but this difference is not statistically different from
zero. Also, there appears to be no change in the fraction of domestic
workers who are employed in full-time positions, nor in the fraction
of workers who are employed as domestic workers in Panel A. Combining this lack of change in employment with the large increases in
mean wages and earnings and the rapid shifts in these wage and
35

CPIX index provided by Statistics South Africa.

In Appendix 2, we use propensity-score reweighting to rule out the possibility that
these shifts in the wage distribution are driven by compositional changes in the type of
domestic workers employed after the law.
36

35

.2

.4

.6

.8

T. Dinkelman, V. Ranchhod / Journal of Development Economics 99 (2012) 2745

4
5
6
7
log monthly earnings for urban domestic workers
Sep01 - before
Sep02 - before
Sep03 - after

Mar02 - before
Mar03 - after
Mar04 - after

Fig. 2. Distribution of log monthly earnings for urban domestic workers. Kernel density plots of log monthly earnings (bandwidth 0.08). Data are from South African Labor Force
Surveys (March 2001March 2004). The vertical line is at the level of the full-time minimum wage (monthly earnings) for urban domestic workers. Each wave of data contains
between 996 and 1260 observations. KolmogorovSmirnov tests of equality of distributions reject at the 5% level for each pairwise comparison of waves in the before and after
periods.

earnings distributions that line up well with the timing of the law, we
have initial, strongly suggestive evidence that the law had a dramatic
impact on the domestic worker sector. Changes in other aspects of the
work relationship also seem important: the fraction of workers with a
job contract increases by 18% in the post period, the fraction with UI
coverage increases by 19 percentage points and the fraction with
any pension contributions increases by 7 percentage points.
All of these comparisons so far rely solely on variation in the application of the law. Particularly for earnings, wage and hours variables

that are likely to uctuate with general economic conditions, it is possible that other factors could inuence some of these outcomes postlaw, thereby confounding the effects of the law. For this reason, we go
beyond the simple pre-post comparison and investigate whether
there are larger wages and hours changes in places where the new
wage oor is more binding.
Fig. 3 provides the basic information we use in our difference-indifferences regression. This gure shows the mean (log) hourly
wage for domestic workers in each wave, for each of nine provinces.

Mean log hourly wage ZAR

1.5

0.5

Wave
WC: -.25

EC: 0.75

NC: 0.57

FS: 0.72

NP: 0.75

Gau: 0.17

MP: 0.49

LIM: 0.79

KZN: 0.57

Fig. 3. Mean hourly domestic worker wages by province over time. Each point represents a province-wave level average hourly wage for domestic workers, for waves before and
after the law. The black vertical dashed lines demarcate the period before (to the left) and the period after (to the right) the law, with the cusp period falling between these lines.
The horizontal dashed line represents the urban full-time minimum wage (ZAR4.1) established in November 2002. All data are from the South African Labor Force Surveys (March
2001March 2004) and means are weighted. Provinces are: Western Cape (WC), Eastern Cape (EC), Northern Cape (NC), Free State (FS), Kwazulu-Natal (KZN), North West Province (NP), Gauteng (Gau), Mpumalanga (MP) and Limpopo (LIM). Province-level wage gap measure calculated prior to the law is shown in the legend for each province: e.g. EC 0.75
means that the median wage was 75% below the minimum wage in the pre-law period.

36

T. Dinkelman, V. Ranchhod / Journal of Development Economics 99 (2012) 2745

Table 2
Log hourly wages of domestic workers: Difference-in-differences.
Excluding cusp wave

All waves
OLS

POST
Pre-law wage gap (WG)
Pre-law wage gap WGj*POST
Age, education, African controls
N

Two-step estimator

OLS

Two-step estimator

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

0.202***
(0.02)
0.858***
(0.04)
0.090**
(0.04)
N
6154

0.193***
(0.02)
0.820***
(0.05)
0.097**
(0.04)
Y
6154

0.201***
(0.03)
0.829***
(0.05)
0.100
(0.07)
N
18

0.189***
(0.03)
0.792***
(0.04)
0.100
(0.06)
Y
18

0.215***
(0.02)
0.902***
(0.04)
0.133***
(0.04)
N
5205

0.203***
(0.02)
0.862***
(0.04)
0.145***
(0.04)
Y
5205

0.217***
(0.03)
0.868***
(0.04)
0.138*
(0.06)
N
18

0.202***
(0.03)
0.827***
(0.04)
0.150**
(0.05)
Y
18

*10%, **5%, and ***1% signicance level. Robust standard errors presented in each column and all regressions are weighted. Each regression contains a constant (coefcient not
shown). In columns 1, 2, 5 and 6, standard errors are Eicker-White, clustered at the province level. Critical values for signicance for two-step estimates are taken from the t distribution, see text for details. Columns 14 include domestic workers in all waves. Columns 58 exclude domestic workers in the September 2002 cusp wave. POST = 1 in March
2003, September 2003 and March 2004; otherwise zero. Pre-law wage gap is the province-level difference in the log(4.1) log(median wage), where 4.1 is the urban full-time
minimum wage introduced in November 2002 and median wage is the median wage of domestic workers in each province across all of the pre-waves (September 2001 and
March 2002).

The dashed vertical lines demarcate the pre-, cusp (September 2002)
and post-periods, while the dashed horizontal line denotes the urban
full-time minimum hourly wage set in November 2002. In the legend
of the gure, we also show the value of each province's pre-law wage
gap measure: for example the Western Cape wage gap measure is
0.25 (meaning that median wages in this province are higher
than the minimum wage by 25% prior to November 2002) while Limpopo province has a wage gap measure of 0.798 (meaning that median wages are almost 80% below the minimum wage before November
2002). Provinces can be grouped into one of three categories: those
provinces paying relatively high wages before the law (Western
Cape and Gauteng), provinces paying middle-range wages before
the law (Mpumalanga, KZN and the Northern Cape) and provinces
paying very low wages before the law (Eastern Cape, Free State,
North West Province and Limpopo).
The graph shows that every province except the Western Cape
had mean hourly wages far below the minimum wage prior to the
law. There is no clear evidence of pre-trends in wages that differ between provinces; in a couple of provinces, mean wages look like
they increase somewhat in the cusp wave of September 2002
(WC, NP and MP). And, although all provinces evidence an increase
in mean wages after September 2002, this increase appears to be
steeper for provinces falling further below the minimum prior to
the law. The gure shows clearly that mean log hourly wage measures are bunching together for provinces further away from the
minimum in the post-law period.
Using the cross-province variation in WGj combined with the timing of the law, we estimate difference-in-differences regressions of
the form in Eq. (1). Results for domestic worker wages are presented
in Table 2. Columns 1, 2, 5 and 6 present OLS estimates with robust
standard errors clustered at the province level. In columns 3, 4, 7
and 8 we present the results from the two-step estimator of Donald
and Lang (2007) along with appropriate standard errors and signicance levels taken from the relevant t-distribution. The rst four columns contain results for the full sample of domestic workers and the
last four columns restrict the sample by excluding domestic workers
in the cusp wave of September 2002these are our preferred estimates. We present results without demographic controls (age, education and African indicator) in each odd-numbered column and results
from regressions that include the demographic controls in each evennumbered column. All regressions are weighted. 37
37
Results from the unweighted regressions for log hourly wages are presented
in Appendix 3 Table 1 for comparison. In these tables, we also present alternate
block-bootstrapped standard errors for the OLS estimates, where the province is
treated as the block.

Across all columns, there is a large, signicant increase in wages in the

post period, of between 18.9 and 21.7%. This reects the information in
Table 1 and in Fig. 3: average wages across all domestic workers increase
signicantly after the law.38 Recall that the pre-law wage gap (WGj) is dened such that the further below the minimum wage the provincial median wage lies, the larger (more positive) this variable is. Not surprisingly,
in places with larger WGj, average wages are signicantly lower in the
pre-period. However, in the POST period, provinces that were further behind are the ones where the wage response is the largest, as indicated in
Fig. 3. The coefcient on POSTt *WGj is large and positive in each specication and signicantly different from zero in both the OLS and two-step estimator results, for the sample that excludes September 2002.39 Focusing
on the last four columns of this table, our estimate of 3 suggests that domestic worker wages increased by 13 to 15% after the law. Or, to take a
particular example: for a worker living in the province with the largest
(demeaned) pre-law wage gap (0.36), the average increase in wages
after the law is about 25% using either the OLS (0.203+0.36*0.145) or
the two-step (0.201+0.36*0.15) results.
In contrast to these large wage effects that appear shortly after the
law, hours of work do not exhibit similar signicant declines in the
POST period.40 Table 3 presents results of the form in Table 2 for usual
weekly hours of work. Across specications, the point estimate on
POSTt is between 0.76 and 1.149 and never statistically signicant. Regardless of the method of estimation, the coefcient on POSTt * WGj is
larger and negative, suggesting that hours may have declined more (between 2.8 and 5.1 hours more) in areas where the initial wage gap
was larger. This is between a 7 and 12% fall in employment on the intensive margin. As an example: for a worker living in a province with the
largest de-meaned pre-law wage gap, the average reduction in hours
of work is about 6% [( 1.11 + 0.36 ( 3.57))/40] using the OLS
results. However, in all specications, we cannot reject that these estimated changes in hours of work are zero; none of the coefcients on
the POSTt * WGj variable is close to being precisely estimated. It is possible that measurement error in reports of hours of work undermines

38
Recall that the wage gap measure is demeaned; so the coefcient on POSTt can be
interpreted as the average change in wages for domestic workers in areas with the average wage gap measure.
39
The relevant critical values from the t distribution for a one-tailed test with 4 degrees of freedom are 2.13 (p b 0.05 signicance) and 1.53 (p b 0.1 signicance). For a
two-tailed test, the relevant critical values are 2.77 (p b 0.05) and 2.13 (p b 0.1). We
use the t distribution because the number of observations in the second step of the estimation is small.
40
Sample size changes across tables as more workers report hours of work information than report monthly earnings.

T. Dinkelman, V. Ranchhod / Journal of Development Economics 99 (2012) 2745

37

Table 3
Usual weekly hours of work of domestic workers: Difference-in-differences.
Excluding cusp wave

All waves
OLS

POST
Pre-law wage gap (WG)
Pre-law wage gap (WG) * POST
Age, education, African controls
N

Two-step estimator

OLS

Two-step estimator

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

0.887
(0.82)
5.151
(3.21)
2.832
(2.17)
N
6876

0.901
(0.81)
3.98
(2.58)
3.309
(2.28)
Y
6876

0.848
(1.58)
5.758*
(2.52)
4.157
(3.79)
N
18

0.761
(1.39)
4.996*
(2.17)
4.283
(3.50)
Y
18

1.149
(1.18)
5.383
(3.84)
3.064
(3.20)
N
5824

1.111
(1.14)
4.1
(3.01)
3.579
(3.23)
Y
5824

0.874
(1.62)
6.675*
(2.94)
5.074
(4.08)
N
18

0.77
(1.40)
5.729*
(2.48)
5.128
(3.70)
Y
18

*10%, **5%, and ***1% signicance level. Robust standard errors presented in each column and all regressions are weighted. Each regression contains a constant (coefcient not
shown). In columns 1, 2, 5 and 6, standard errors are Eicker-White, clustered at the province level. Critical values for signicance for two-step estimates are taken from the t distribution, see text for details. Columns 14 include domestic workers in all waves. Columns 58 exclude domestic workers in the September 2002 cusp wave. POST = 1 in March
2003, September 2003 and March 2004; otherwise zero. Pre-law wage gap is the province-level difference in the log(4.1) log(median wage), where 4.1 is the urban fulltime minimum wage introduced in November 2002 and median wage is the median wage of domestic workers in each province across all of the pre-waves (September 2001 and March
2002).

Table 4
Probability of working as a domestic worker: Difference-in-differences.
Excluding cusp wave

All waves
OLS

POST
Pre-law wage gap (WG)
Pre-law wage gap (WG)*POST
Age, education, African controls
N

Two-step estimator

OLS

(1)

(2)

(3)

(4)

(5)

(6)

Two-step estimator
(7)

(8)

0.004
(0.01)
0.0409***
(0.01)
0.004
(0.02)
N
52,739

0.003
(0.01)
0.014
(0.01)
0.002
(0.02)
Y
52,739

0.012
(0.01)
0.0594**
(0.02)
0.024
(0.03)
N
18

0.005
(0.01)
0.008
(0.02)
0.017
(0.03)
Y
18

0.010
(0.01)
0.0386**
(0.02)
0.002
(0.03)
N
44,005

0.002
(0.01)
0.0173**
(0.01)
0.004
(0.03)
Y
44,005

0.014
(0.01)
0.0610**
(0.02)
0.025
(0.03)
N
18

0.007
(0.01)
0.007
(0.02)
0.018
(0.03)
Y
18

*10%, **5%, and ***1% signicance level. Robust standard errors presented in each column and all regressions are weighted. Each regression contains a constant (coefcient not
shown). In columns 1, 2, 5 and 6, standard errors are Eicker-White standard errors, clustered at the province level. Critical values for signicance of two-step estimates are
taken from the t-distribution, see text for details. Sample in columns 14 include employed and unemployed women aged 2059 (searching unemployed) who have no more
than high school; sample in columns 58 excludes all of the individuals in this group who appear in the September 2002 cusp wave.

our ability to precisely estimate the effect of the law on hours of work.
Nevertheless, we nd no strong statistical evidence that employers adjusted labor demand on the intensive margin in order to afford the massive increase in wages that are evident in the data.41
There is also no evidence that adjustment occurred on the extensive
margin. If domestic workers lost jobs as a result of the law, we should
see different probabilities of low-skilled African and colored women
being employed as domestic workers POST-law. Table 4 presents
difference-in-differences results for the binary outcome Does the individual work as a domestic worker? The sample includes domestic
workers and demographically similar women who are working or
searching for work. Dening the sample in this way allows for the possibility that domestic workers may switch to other jobs or lose jobs altogether in the POST period. None of the estimated coefcients on the
POSTt or POSTt *WGj variables is large, or statistically signicant, under
any specication.42

41
Results for the difference-in-differences coefcient in the wage and hours regressions are the same if we also include a full set of province xed effects to control for
level differences in wages across provinces.
42
Other types of extensive margin adjustments (of type rather than number of domestic workers) may have altered the composition of domestic workers and contributed to observed earnings shifts. For example, in the POST-period, employers might try
harder to select higher quality workers once the law is in place and as a result pay these higher quality domestic workers more. That these changes drive the results we see
is ruled out in our propensity-score reweighting exercise in Appendix 2.

4.2. Checking for threats to validity

There are three main concerns with the difference-in-difference
strategy that we use to identify the causal impact of the new law on
wages and employment. One concern regards accurate inference:
with such a small number of effective units driving the main variation in the intensity of the law, we need to be cautious that we are
not over-rejecting a null of zero effect. The results from the twostep estimator in the previous section address this concern. The
other concerns relate to direct threats to the validity of our causal
estimates.
The primary threat to validity is that provinces with different prelaw WGj measures may also experience differential economic trends
in the POSTt period or contemporaneous shocks, which could account
for our estimates of 3. Having more prior years of data generally
helps to rule out a differential trends explanation; however, as mentioned before, we are limited in the amount of pre-law data we can
use to investigate this. Instead, we tackle this issue by understanding
more about where the underlying wage gap variation is coming from,
and by implementing a placebo test.
First, it is useful to consider the type of differential trend that
could confound the wage and employment results of the previous
section. Recall from Fig. 3 that, when ordering provinces on WGj, the
Western Cape, Gauteng, Mpumalanga and KZN emerge as the highest
paying provinces prior to the law. Three of these four provinces comprise the economic centers of South Africa, the provinces that contain the three largest cities of Johannesburg, Durban and Cape
Town. These are areas where labor demand in general is much

T. Dinkelman, V. Ranchhod / Journal of Development Economics 99 (2012) 2745

ZAR Millions in 2005 prices

100
300
400
200

500

38

2000

2001
WC
NWP

2002
year
EC
Gau

2003
NC
MP

FS
Lim

2004
KZN

Fig. 4. Real GDP by province and year. Data are from National Treasury. Provinces are: Western Cape (WC), Eastern Cape (EC), Northern Cape (NC), Free State (FS), Kwazulu-Natal
(KZN), North West Province (NWP), Gauteng (Gau), Mpumalanga (MP) and Limpopo (LIM).

stronger than in other parts of the country; they jointly account for
close to two-thirds of the country's GDP. We can see this in Fig. 4
which presents annual real GDP in millions of ZAR at the province
level for each year from 2000 to 2004. The level of output produced
by Gauteng, KZN and the WC clearly dominate the contributions of
the other provinces. For our positive estimate of 3 to be driven by
differential labor demand, we would need to see strong improvements in the provincial economies with the largest wage gap values
(i.e. Limpopo, the Eastern Cape, the North West Province, the Free
State, and the Northern Cape), or strong declines in the provincial
economies with the smallest wage gap values. This is not, in fact,

what we see in Fig. 4: over time the small wage gap provinces experience some trend growth in annual GDP, while provinces with larger
wage gaps show no signs of strong positive trends in GDP growth.
Strong labor demand trends that differ across provinces are therefore
unlikely to explain the large wage effects we nd in Table 2, and the
lack of signicant negative employment effects in Table 3.
We can more formally rule out the possibility that high wage gap
provinces experienced large general shocks to their economies at the
same time as the minimum wage law came into effect using a placebo
test. We ask: do wages rise (hours of work fall) in high wage gap relative to low wage gap provinces, for workers who are similar to

Table 5
Difference-in-differences for male manufacturing workers.
Log hourly wages

Weekly hours of work

OLS
(1)

(2)

Two-step estimator

OLS

(3)

(5)

(6)

(7)

(8)

0.913***
(0.21)
2.208**
(0.95)
1.366**
(0.59)
4824

0.880***
(0.22)
1.670*
(0.82)
1.208*
(0.55)
4824

0.755
(0.593)
2.461*
(1.101)
0.795
(1.438)
18

0.736
(0.554)
2.013
(0.957)
0.744
(1.271)
18

0.999***
(0.269)
0.029
(3.695)
0.275
(2.553)
4824

0.684
(0.825)
1.302
(6.085)
0.877
(8.197)
18

0.671
(0.740)
1.992
(5.308)
0.591
(7.211)
18

(4)

A: Placebo test using manufacturing worker wages/hours and the pre-law domestic worker wage gap
POST
0.103***
0.0956***
0.089
0.09
(0.01)
(0.01)
(0.05)
(0.05)
Pre-law wage gap
0.263**
0.240**
0.242**
0.203*
(0.08)
(0.09)
(0.09)
(0.08)
Domestic wage gap*POST
0.042
0.042
0.002
0.003
(0.04)
(0.04)
(0.13)
(0.13)
N
3631
3631
18
18

B: Robustness check for mean reversion using manufacturing worker wages/hours and the pre-law manufacturing worker wage gap
POST
0.107***
0.100***
0.0888**
0.0894**
1.050***
(0.015)
(0.014)
(0.033)
(0.037)
(0.268)
Pre-law wage gap
1.038***
0.883**
0.981***
0.811***
3.321
(0.238)
(0.290)
(0.245)
(0.183)
(4.881)
Manufacturing wage gap * POST
0.097
0.137
0.146
0.159
0.993
(0.200)
(0.156)
(0.339)
(0.301)
(2.567)
N
3631
3631
18
18
4824

Two-step estimator

*10%, **5%, and ***1% signicance level. Robust standard errors presented in each column and all regressions are weighted. Each regresssion contains a constant (coefcient not
shown). In columns 1, 2, 5 and 6, standard errors are Eicker-White, clustered at the province level. Critical values for signicance of between estimates and for two-step estimates
are taken from the t-distribution, see text for details. Sample includes all African and colored male workers of relevant age and education level employed in manufacturing, in all
waves excluding the cusp wave September 2002. POST = 1 in March 2003, September 2003 and March 2004; otherwise zero.
Panel A: Pre-law wage gap is the province-level difference in the log(4.1) log(median wage), where 4.1 is the urban full-time minimum wage introduced in November 2002 and
median wage is the median wage of domestic workers in each province across all of the pre-waves (September 2001 and March 2002). Panel B: Pre-law wage gap is the provincelevel difference in the log(4.1) log(median manufacturing worker wage), where 4.1 is the urban full-time domestic worker minimum wage introduced in November 2002. Median manufacturing worker wage is dened at the province level across all of the pre-waves (September 2001 and March 2002).

T. Dinkelman, V. Ranchhod / Journal of Development Economics 99 (2012) 2745

domestic workers but who are not affected by this law? If the answer
is yes, we might be concerned that our estimate of 3 is not really
picking up the impact of the minimum wage law.
Table 5 Panel A presents the results of this placebo experiment.
We implement our main estimating equation in (1) for male
manufacturing workers who are employed in urban areas, between
age 20 and 59 (inclusive), are African or colored and who have no
more than a high school level of education. The rst four columns of
the table present OLS and two-step estimator results for log hourly
wages, and the nal four columns present OLS and two-step estimator results for weekly hours of work. For brevity, we only present
the results that exclude the September 2002 cusp wave.
For this sample of manufacturing workers, wages are 9 to 10%
higher in the POST period, although this change is not signicant
once we take into account the grouped nature of the data (columns
3 and 4). Importantly, the change in wages for manufacturing
workers in high wage gap areas does not seem to be signicantly
higher in the POST period. The coefcients on WGj * POSTt are small
and statistically insignicant. Hours of work are higher in the post period in high wage gap areas (columns 5 and 6), although these are
again, not signicantly different from zero once we take into account
the grouped nature of the error term in Eq. (1) in the presence of a
small number of groups. These estimates together with Fig. 4 provide
some evidence that our difference-in-differences estimates of the
wage effects of the law for domestic workers in Table 2 do reect
the impact of the law and not simply positive wage trends or contemporaneous positive shocks in high wage gap provinces.
A nal concern posing a threat to the validity of our estimates of
3 relates to mean reversion in wages: wages might rise in more intensely treated province simply because of mean reverting measurement error. To some extent, using province-level aggregate data
insulates us from the worst forms of thisthe problem would be
more severe if our wage gap measure was computed at the level of
a smaller geographic entity with fewer observations per unit, and
worse still if we were using individual level data to dene this
treatment measure. We can, however, provide some evidence
against mean reversion explaining our results even at the provincial
level, using the sample of male manufacturing workers. If wages
are strongly mean-reverting, then this should show up as low
paid workers (based on a measure of manufacturing wages) being
paid more after the law, even for workers who are unaffected by
the law.
To implement this test, we regress the wages of male manufacturing workers on a POSTt indicator, a different wage gap measure that
captures the difference between the domestic worker minimum in
2002 (an arbitrary benchmark) and the median manufacturing worker's wage at the province level before 2002, and their interaction. Interestingly, the correlation between this wage gap measure and our
original wage gap measure for domestic workers is relatively low at
0.4. Because of this, the mean reversion test is quite a different test
than our placebo experiment.
The results of testing the mean reversion hypothesis for log male
manufacturing worker wages are shown in Panel B of Table 5. Focussing
on the coefcients on the interaction terms, there is no clear indication
of mean reversion in the data. None of the coefcients are close to statistically signicant at conventional levels. We take this as supportive
evidence that mean reversion in wages (or hours of work) for domestic
workers cannot account for the main wage results.
4.3. Testing for partial compliance
In this section, we provide suggestive evidence that some of the
wage increases that occur after the law are only in partial compliance
with the minimum. This is important to show, because it gives us
some sense of how this labor market operates. It suggests that the
effects of the law are not simply driven by a subset of employer-types

39

Table 6
Testing for partial compliance in response to the law.
Change in predicted probability Controlling only
for POST
of domestic worker reporting
nominal wages in the
1 = 0.9 wmin
near-compliant region

Controlling for
POST, WG and
POST * WG
1 = 0.9 wmin

Pr(w* = 1|POST)
Pr(w* = 1|PRE)

No
controls

All
controls

No
controls

All
controls

(1)

(2)

(3)

(4)

0.007***

0.007***

0.012***

0.001***

(0.001)

(0.001)
0.047***

(0.000)
0.002***

(0.000)

(0.004)
5205

(0.000)
5205

Pr(w* = 1|POST * WG)

Pr(w* = 1|PRE * WG)
N workers

5205

5205

*10%, **5%, and ***1% signicance level. Coefcients are predicted changes in the
probability of earning a wage in the near compliant bin where the predictions are
generated from estimated coefcients from an ordered probit model (using Stata's
non-linear prediction command). The three groups in the ordered probit are: noncompliant (w = 0), near-compliant (w = 1) and compliant (w = 2) and we dened
near compliant as having a wage 10% below the minimum or less. In columns (3)
and (4), we compute the predictions taking into account the impact of an interacted
variable in a non-linear model. Sample excludes the cusp wave (September 2002).
Each specication includes an indicator for whether the observation is captured PRE
or POST law. Columns (2) and (4) present results from models which control for age,
race, years of education. Columns (3) and (4) present results from the specication
that controls for a POST indicator, a measure of the log wage gap in the province before
the law was in effect, and the interaction of POST with wage gap (WG) (i.e. the difference in differences specication).

who want to abide by the letter of the law; rather some employers
are voluntarily choosing whether and by how much to comply with
the law.
Table 6 presents results from the ordered probit model we estimate
for the outcome variable that classies a worker's wage into a range
below, near or above the minimum in each wave, where near is dened
as being paid 9099% of the minimum wage. 43 Recall from the discussion above (and Appendix 1), the idea of the test is that if the probability
of a worker reporting a wage in the non-compliant bin falls by more
than the probability of a worker earning a wage in the compliant bin
rises, then some workers are earning more than they were before, but
not enough more to bring them into compliance with the law. This is
equivalent to testing whether the probability of a worker being in the
near compliant bin rises after the law, and the test is informative as
long as there are no large employment reductions in response to the
law, which we showed in Tables 3 and 4.
Since there are a range of coefcients we could report from the ordered probit, we isolate the marginal effects for the POSTt indicator
(all columns) and the interaction term (last two columns) on the
probability of being classied in the near compliant bin. In the
rst two columns, our specication excludes the POSTt WGj control;
the interaction term is included in the specication underlying the estimates in the second columns. We compute these marginal effects
using non-linear prediction methods.
To interpret results, consider the coefcients in column (1). In the
POST period, the probability of a worker reporting hourly wages in the
near compliant range increases by a signicant 0.7 percentage points:
more workers are squeezing into the narrow band near the minimum.
This result looks across all workers and compares the pre- to the postperiod. Columns (3) and (4) show that this shift in the direction of the
wage oor is more pronounced in higher wage gap provinces, after

43
Results are similar whether we choose the cut-off for being near the minimum as
20% or less than the minimum.

40

T. Dinkelman, V. Ranchhod / Journal of Development Economics 99 (2012) 2745

the law. Without demographic controls, there is a 4.7 percentage point

increase in the probability of a worker earning a wage close to the minimum; this change falls to 0.2 percentage points when we add demographic controls, yet is still statistically different from zero.
We interpret these results as indicating partial compliance with
the law at the lower parts of the wage distribution. The evidence on
increasing dispersion (variance) in wages and earnings in Table 1
and the graphical evidence on increases in mean log hourly wages
for domestic workers even in the Western Cape (Fig. 3)the only
province where the mean wage was initially above the minimum
point toward an employer response at levels much higher in the distribution. Together, these results suggest that some employers are
responding to the law in ways different to those predicted by conventional compliance models. In the next section, we present a nal test
that tries to rule out the possibility that employer responses are driven by a desire to avoid penalties associated with non-compliance.
4.4. Testing for compliance related to probability of audit
To learn more about how much employer behavior may be driven by
the threat of external sanctions, we investigate the responsiveness of
wages to the presence of a local LC. While this is not the only aspect of
a local labor market that could increase the likelihood of being caught
for non-compliance, it is a plausible feature that distinguishes markets
and it is feasible to obtain data on the location of these ofces. Since
dealing with worker complaints is one of the three main responsibilities
of these LCs, living nearby an LC should capture a lower cost of complaint for domestic workers, regardless of whether or not these workers
actually use these labor centers in equilibrium.44 Hence, having an LC
nearby should increase the actual probability of audit as well as employer beliefs about their likelihood of being monitored.
We tracked down the physical addresses of these LCs and matched
each one to a unique magisterial district in the LFS, a geographic unit
that is more disaggregated than the province. Since the LCs were primarily established to serve formal sector workers in the rest of the
economy, it is unlikely that their location is endogenous to the prevalence of domestic worker employers or to the presence of noncompliant employers of domestic workers; however, the LCs are likely
to be over-represented in areas with more economic activity in the formal sector. Across all waves, 75% of domestic workers live in magisterial
districts where there is at least one LC.
In Table 7, we estimate wage regressions of the form in Table 2,
now including a control for whether the domestic worker lives in a
magisterial district with an LC, the interaction of this indicator with
POSTt, with WGj and the triple interaction POSTt*WGj*LCijt. We estimate Eq. (1) again using OLS (and clustered standard errors) as well
as using the two-step estimator and present results for the sample excluding September 2002, for brevity. Interestingly, the OLS results indicate that domestic worker wages are about 12% higher in areas that
have an LC (in columns (1) and (2)) and that LCs tend to offset the effect of working in a province where the median wage in the preperiod is below the minimum (coefcient on LC * WG). However, neither of these differences are evident once we estimate the model
using the two-step procedure. Furthermore, there seems to be no indication that having an LC in one's local labor market increases the
impact of the minimum wage for domestic workers; employers who
raise wages are not doing so differentially in areas where the

44
LCs also pay out UIF claims and deal with general enquiries, see http://www.
labour.gov.za/contacts/Labour%20Centres/labour-centres-gauteng-south for details.
The Labor Centers are also, practically, where labor inspectors actually work. These inspectors are responsible for enforcing sectoral determinations across all industries
(there were 57 inspectors per 1 million workers in 2007 (Bhorat et al., 2010)) and so
are stretched a bit thin, reducing the overall audit probability for any one industry.
However, as noted by Bhorat et al. (2010), inspections by these labor centers are generally triggered by clients and investigated on a case-by-case basis.

Table 7
Wage responses in areas with high versus low costs of complaint: Difference-indifferences.
OLS
(1)
POST

0.258***
(0.02)
Pre-law wage gap (WG)
1.008***
(0.04)
Pre-law wage gap (WG) * POST
0.180**
(0.06)
Labor Center
0.122***
(0.02)
Labor Center * POST
0.0568
(0.03)
Labor Center * WG
0.175***
(0.04)
Labor Center * WG * POST
0.079
(0.07)
Controls for age, education, African? N
N
5205

Two-step
estimator
(2)

(3)

(4)

0.257***
(0.02)
0.979***
(0.06)
0.196***
(0.05)
0.121***
(0.02)
0.0716**
(0.03)
0.193***
(0.05)
0.0884
(0.07)
Y
5205

0.155
(0.23)
0.483
(0.30)
0.804
(0.61)
0.12
(0.20)
0.527
(0.32)
0.647
(0.48)
0.807
(0.89)
N
36

0.164
(0.23)
0.422
(0.30)
0.812
(0.64)
0.058
(0.20)
0.52
(0.32)
0.672
(0.48)
0.798
(0.93)
Y
36

*10%, **5%, and ***1% signicance level. Robust standard errors presented in each column
and all regressions are weighted. Each regression contains a constant (coefcient not
shown). In columns 1 and 2 standard errors are Eicker-White, clustered at the
province level. Critical values for signicance for two-step estimates are taken from the
t-distribution, see text for details. Sample includes domestic workers in all waves excluding the September 2002 cusp wave. POST= 1 in March 2003, September 2003 and
March 2004; otherwise zero. Labor Center= 1 if the domestic workers in a magisterial
district that contains at least one Labor Center (LC).
Pre-law wage gap is the province-level difference in the log(4.1) log(median wage),
where 4.1 is the urban full-time minimum wage introduced in November 2002 and
median wage is the median wage of domestic workers in each province across all of
the pre-waves (September 2001 and March 2002).

likelihood of being caught for non-compliance is higher. Although it

is never possible to precisely estimate the coefcient on LC * WG * POST,
the coefcient on this triple-difference term is negative, pointing in
the wrong direction for the mechanism that has employers responding more in areas with a higher probability of audit.
4.5. Contract coverage and related employment benets
Since the sectoral determination for domestic workers mandated
formal labor contracts and some additional rights and benets, it is likely that the law also impacts the conditions of work for domestic
workers, over and above any direct effects on wages. We already see
strong evidence of this in Table 1, with the large increases in contract
coverage, UI and pension benets after November 2002. In this section,
we investigate whether workers originally paid below the minimum
wage are more likely to also see improvements in their employment
benets and legal rights (in terms of formal contract coverage).
In Table 8, we present difference-in-differences comparisons for
whether a worker has a formal written contract of employment,
whether the employer contributes to an unemployment insurance
fund, and whether the employer contributes to a pension fund. The
coefcient on the POST indicator reects the large differences in
pre-post means that were seen in Table 1. The fraction of workers
with a formal job contract rises between 17 and 20 percentage points,
the fraction of workers with UI coverage rises by 18 to 20 percentage
points and the fraction enjoying pension contributions rises by 67
percentage points. Notice that workers in high wage gap provinces
are not likely to experience larger improvements in these conditions
of work relative to low wage gap provinces after the law. Although
the interaction terms for the UI outcomes are negative in the OLS regressions, none of the interaction terms is statistically signicant in
any of the two-step estimation results.
While it may not be surprising that improvements in employment
benets and rights are unrelated to initial levels of non-compliance
with the wage part of the minimum wage law, it is striking that

T. Dinkelman, V. Ranchhod / Journal of Development Economics 99 (2012) 2745

Table 8
Improvements in domestic worker employment rights and benets: Differenceindifferences.
OLS
(1)

Two-step estimator
(2)

(3)

(4)

0.200***
(0.03)
0.100**
(0.03)
0.07

0.196***
(0.03)
0.075**
(0.02)
0.07

(0.07)
N

(0.06)
Y

18

18

0.205***
(0.02)
0.0387***
(0.01)
0.054

0.202***
(0.02)
0.0217*
(0.01)
0.052

(0.05)
N

(0.04)
Y

18

18

0.0634***
(0.01)
0.0500**
(0.02)
0.040

0.0707***
(0.01)
0.0424***
(0.01)
0.037

0.0680***
(0.01)
0.508**
(0.01)
0.037

(0.03)
Y

(0.03)
N

(0.03)
Y

5815

18

18

A: Worker has a formal job contract with employer

POST
0.173***
0.170***
(0.02)
(0.02)
Pre-law wage gap (WG)
0.101**
0.073**
(0.03)
(0.03)
Pre-law wage gap
0.07
0.06
WG * POST
(0.05)
(0.05)
Age, education,
N
Y
African controls
N
5743
5743
B: Employer makes UIF contributions on behalf of worker
POST
0.184***
0.181***
(0.01)
(0.01)
Pre-law wage gap (WG)
0.0372**
0.017
(0.01)
(0.01)
Pre-law wage gap
0.0731*
0.0674*
WG * POST
(0.03)
(0.03)
Age, education,
N
Y
African controls
N
5817
5817
C: Employer contributes to a pension fund
POST
0.0653***
(0.01)
Pre-law wage gap (WG)
0.0439***
(0.01)
Pre-law wage gap
0.039
WG * POST
(0.03)
Age, education,
N
African controls
N
5815

*10%, **5%, and ***1% signicance level. Robust standard errors presented in each
column and all regressions are weighted. All outcomes are binary, and each
regression contains a constant (coefcient not shown). In columns 1 and 2, standard
errors are Eicker-White, clustered at the province level. Critical values for signicance
for two-step estimates are taken from the t-distribution, see text for details. Sample excludes domestic workers in the September 2002 cusp wave. POST = 1 in March 2003,
September 2003 and March 2004; otherwise zero. Pre-law wage gap is the provincelevel difference in the log(4.1) log(median wage), where 4.1 is the urban full-time
minimum wage introduced in November 2002 and median wage is the median wage
of domestic workers in each province across all of the pre-waves (September 2001
and March 2002).

after the new Sectoral Determination comes into place, workers

across the board are more likely to have a formal job contract, UI
and pension benets. Considering the low baseline contract coverage
rates (10%), UI coverage (2%) and pension coverage (3%), this more
than doubling of contract and pension coverage rates, and tripling
of UI coverage, in the year after the law is enacted is remarkable. It
is unlikely that anything other than the introduction of the minimum
wage law could have had such immediate impacts on the conditions
of work for these workers. Importantly, nothing in the Sectoral Determination made pension contributions mandatory for domestic worker employers. Although pension contributions are stipulated in the
general Basic Conditions of Employment Act covering all workers, as
we noted earlier, the BCEA was seldom adhered to in the informal
sector. The large increases in the fraction of domestic workers enjoying pension contributions after the law suggests that the introduction
of the minimum wage law may have been a catalyst for employers to
start adhering to other aspects of general labor legislation.
We view this last set of results as particularly important, because
they indicate that with the introduction of sector-specic protections
for a previously informal industry, there is a shift in the employment
relationship towards the more formal. Workers gain legal protections

41

and real benets previously denied, and so the new regulation appears to have initiated the formalization of the industry. This process
could have far-reaching consequences for the nature of domestic
work in South Africa.

5. Discussion and conclusions

The introduction of a new minimum wage law for an informal
market presents a unique opportunity to examine important issues
around responses to legal wage oors. It also allows us a window
onto how informal labor markets operate and the conditions under
which they might become more formal.
Although conditions in and characteristics of the domestic worker
industry in South Africa were stable before the introduction of a minimum wage, the difference-in-differences results clearly indicate that
domestic workers who worked in areas where the pre-law median
wage was below the minimum (i.e. where the new law had more
bite) experienced large increases in wages in the POST-period: on
average, wages rose 13 to 15% for workers in provinces with the
mean wage gap. Despite the absence of full compliance and a sharp
wage spike at the minimum, we nd evidence of a strong wage response to the law, and little statistical evidence of work reductions
on the intensive or extensive margins. The main purpose of our
paper has been to document these changes and to provide evidence
that labor market regulation in an informal sector of considerable importance can have a real and immediate impact, even with very limited monitoring and enforcement. The results points to a highly
inelastic demand for domestic workers among private households,
at least in the short run. Our nal set of results also point to signicant
improvements in the conditions of work across the entire domestic
work sector. They are particularly useful in showing that real formalization of the industry is a likely consequence of the new laws.
Given that issues of compliance are not often at the center of empirical work on minimum wages, it is worthwhile considering how our results should be interpreted. One possible reason that the law had an
effect is that some fraction of employers erroneously believed that the
government was going to enforce the new minimum and penalize
non-compliers. A difculty with this explanation is that it cannot account for wage or earnings increases for domestic workers already
paid more than the minimum before the law. Although we do not
have a panel data set of workers to test this directly, the fact that the variance of the wage and earnings distribution increases rather than compresses after the law suggests that some employers increased wages in
excess of the minimum, despite already being compliant with the law.
A second reason that a new and largely unenforced law may have
been effective relates to models of fairness in wage-setting. A theoretical literature in labor economics posits that the notion of a fair wage
is important in incentivizing workers to provide high effort in tasks
for which effort is unobservable (Akerlof, 1982, 1984 and Akerlof
and Yellen, 1988). These models are difcult to test empirically,
since dening and measuring a fair wage, or a reference wage, is
tricky in practice. 45 Experimental studies have separately established
the importance of gift exchange and fairness intentions in employer
employee relationships. 46 In the case of the domestic worker industry
in South Africa, it is plausible that the announcement of a wage oor
45

Mas (2006) is a notable exception.

See for example Brandts and Charness (2004), who show that a minimum wage
may undermine the efciencies that gift-exchange can achieve, Falk et al (2000) discuss how fairness intentions matter for behavior and Fehr et al (1997) model how reciprocity in response to fair behavior expands the set of enforceable contracts that can
be sustained in an economy. Konow (2003) reviews the large body of experimental evidence on fairness. Falk et al. (2005) provide a model and experimental evidence that
the announcement of a minimum wage raises workers' reservation wages in a persistent way. Reservation wages remain high even after the wage oor is removed, suggesting that policy interventions may have direct and indirect effects on behavior by
altering the meaning of a fair transaction and by creating entitlement effects.
46

42

T. Dinkelman, V. Ranchhod / Journal of Development Economics 99 (2012) 2745

dened, or re-dened, what the fair wage was and set in motion voluntary employer responses, even though the two traditional channels
for encouraging complianceenforcement and penaltieswere largely closed off. 47 What we take from our analysis is that in the initial
stages of labor market formalization, governments may need to accept partial compliance with new legislation in order to bring about
real changes in outcomes without signicant disemployment. 48
Even with very limited enforcement, such sector-specic legislation
can move the market towards a more formal setting, if, for example,
it increases contract coverage, as in the South African case.
We emphasize that our conclusions are valid for the domestic
worker industry, which is one example of an informal sector. The dynamics of the employment relationship between a single employer
and a single employee no doubt condition the response to the law
and so may only be relevant in some types of informal enterprises.
However, this characterization of the informal sector may not be too
far from a description of the typical small-scale rm that generates
much informal sector employment in developing countries. 49 Finally,
our analysis is only relevant for the short-run effects of a new minimum wage policy. While we nd clear effects in the 16 months
after the law, as employers face annual increases in this minimum
(often above ination in the case of South Africa) and as workers
sort across space in response to these new protections, the wage,
earnings and employment effects of the policy may change as the sector itself becomes more formal over time.
Appendix 1. Testing for evidence of incomplete wage adjustment
Consider Fig. A.1 below which shows the wage distribution F(.) of
domestic worker wages in the period before and after the law. FAfter is
shifted to the right of FBefore and the solid black vertical line at 2 represents the minimum wage level. Dene 1 as 90% of the minimum
wage. Then, classify all worker wages into one of three groups:
wages below 1 are non-compliant, wages between 1 and 2 are
nearly compliant, and wages at or above 2 are compliant.
We would like to know whether the shift in the wage distribution in
response to the law is consistent with incomplete adjustment up to
compliant levels. That is, are wages shifting up out of the bottom part
of the distribution, but not enough to get most workers up to or over
the minimum threshold? Some shifts in the distribution are consistent
with this idea and some are not. In example 1 Appendix 1 Table 1, the
rst set of changes in the fraction of workers in each region would be
consistent with complete adjustment for some workers up to compliance: 10% of workers in each of the non-compliant and nearcompliant regions leave these regions; and 20% of workers join the
compliant region. The second example is, however, not consistent
with full compliance for workers who experience wage changes: 20%
of workers leave the non-compliant region, and the near-compliant region grows by 10% of workers as does the compliant region. While we
cannot say which exact workers are getting wage increase, as long as
jobs are not being lost as a result of the law change, then such shifts in
the overall distribution reect partial adjustments in response to the
law.
Intuitively, we would like to know whether or not the fall in the
fraction of domestic workers in the non-compliant region is fully offset by the increase in fraction of workers in the compliant region. We
do this by implementing an ordered probit model and testing a
47
Rebitzer and Taylor (1995) present a theoretical model of a labor market in which
employment increases when the minimum wage increases, as a result of an efciency
wage. However, they assume compliance with the law in their model. They focus on
showing how an efciency wage model can predict an increase in labor demand following a minimum wage change.
48
See Basu et al. (2007) for a theoretical discussion of why non-compliance with a
minimum wage could be acceptable.
49
Banerjee and Duo (2007) show that small-scale rms and entrepreneurial activities are an important source of income for the poor.

specic hypothesis. To describe the appropriate hypothesis, consider

the following linear model of the log of the individual worker's

Appendix 1 Table 1
Example 1: Fraction of distribution in each region
Non-compliant region
Before
After

Before
After

Near compliant

0.3
0.3
0.2
0.2
Example 2: Fraction of distribution in each region
Non-compliant region
Near compliant
0.3
0.3
0.1
0.4

Compliant
0.4
0.6
Compliant
0.4
0.5

wage (wi) (ignoring covariates for now) where
i N(0, 1): wi = 0 +

t1 +
i, where t = 1 if the worker is observed in the after period and
t = 0 if observed in the before period. We dene an ordered categorical
variable, wi that captures the region that each wage falls into:


wi 0


wi 1

if

wi 1

if

1 bwi 2

A test for partial compliance



wi 2

if

wi > 2

Fig. A.1: A test for partial compliance.

Then, the probability that any worker's wage falls into a particular
bin can be described as:
 

Pr wi 0 Pr wi 1
Pr 0 t1
i 1 1 0 t1

 

Pr wi 1 Pr 1 b0 t1
i 2 2 0 t1
1 0 t1

 

Pr wi 2 Pr 0 t1
i 2 1Pr 0 t1
i 2
12 0 t1

T. Dinkelman, V. Ranchhod / Journal of Development Economics 99 (2012) 2745

The hypothesis we would like to test is the following:

H0 :

Pr wi 2 Pr wi 0

0
t
t

HA :

Pr wi 2 Pr wi 0

b0
t
t

Let the parameter combination described in the hypothesis statement be . If we reject 0, this provides evidence consistent with
some domestic workers getting wage increases that are up to a level
less than the minimum, i.e. up to a new but still non-compliant level
in region 2 1.
Implementing this test is straightforward. To see this, note that we
can write each of the partial derivatives in this way:
Pr wi 2
1 2 0 t1
t

Pr wi

1 1 0 t1

In words, we want to know whether the probability of being in the

near compliant region rose in the after period, relative to before the
law was in place.
We can also extend the test to a framework that includes a measure of
the intensity of the law. Suppose that wages are given by wi =0 +t1 +
WGj2 +WGjt3 +
i, where WGj is the median wage gap for domestic
workers in the local labor market during the period before the law. This
implies that we can write the partial derivative with respect to t as:

Pr wi 1

1 WGj t3 2 0 t1 1 0 t1  15
t


Our test will involve examining the signs and signicance of the
marginal effect of t, at the mean level of the pre-law wage gap. We compute the marginal effects and associated standard errors of the POST and
POSTt * WGj measures using non-linear prediction methods in Stata.
Appendix 2. Checking whether composition effects explain the
shift in wage distributions

10

This allows us to redene as:



Pr wi 2 Pr wi 0

t
t
1 2 0 t1 1 0 t1 

43

11

Finally, note that:

Pr wi 1
1 2 0 t1 1 0 t1 
t

12

We use this last partial derivative to re-state the null and alternative hypotheses:
H 0 : 0

13

H A : > 0

14

We use a simple propensity score re-weighting technique (as in

DiNardo et al. (1996)) to show that the shift in the distribution of observable characteristics (Appendix 2, Fig. 1) for workers accounts for a
small fraction of the actual shift in wages. We check whether the distribution of observable characteristics of domestic workers changes significantly over the period by estimating a probit model of the probability of
being a domestic worker in the PRE-period (yi = 0) or the POST period
(yi = 1) and plotting the distribution of predicted probabilities for
each period in Appendix 2 Fig. 1. There is substantial overlap in the propensity scores in the two periods but also a noticeable rightwards shift
in the distribution of scores POST-law. To check whether these changes
can account for the large shifts in earnings we see after the law, we
apply a propensity score weight to the earnings data of observations
in the pre-period and graph three kernel density plots of the distribution of earnings reported by workers: the pre-law distribution, the
POST distribution and the POST distribution re-weighted for the distribution of characteristics observed in the pre-period (as in DiNardo et
al., 1996). Appendix 2 Fig. 2 shows that re-weighting in this way does
not eliminate the shift in wages from the PRE to the POST-period.

Appendix 2

Overlap in predicted propensity scores: Probability of being a domestic worker

10

Before (Sep01, Mar02, Sep02) and After (Mar03, Sep03 and Mar04)

.3

.4

.5
predicted p-scores from a probit
before

.6

after

Fig. 1. Shift in observable characteristics of urban domestic workers.

.7

44

T. Dinkelman, V. Ranchhod / Journal of Development Economics 99 (2012) 2745

.2

.4

.6

.8

Appendix 2

-2

loghourlywage
solid: pre-law
dot: pre-law, weighted to look like post-law

dash: post-law

Fig. 2. Reweighted wage distributions.

Appendix 3
Appendix 3 Table 1: Log hourly wages of domestic workers: Difference-in-differences, unweighted.
Excluding cusp wave

Two-step estimator
OLS

POST

Pre-law wage gap (WG)

Pre-law wage gap (WG)*POST

Controls for age, education, African?

N

Two-step estimator

OLS

(1)

(2)

(3)

(4)

(5)

(6)

Two-step estimator
(7)

(8)

0.198***
(0.01)
[0.015]
0.829***
(0.05)
[0.167]
0.109***
(0.03)
[0.052]
N
6154

0.188***
(0.01)
[0.017]
0.808***
(0.05)
[0.118]
0.111***
(0.03)
[0.066]
Y
6154

0.194***
(0.03)

0.183***
(0.03)

0.197***
(0.03)

0.787***
(0.04)

0.848***
(0.04)

0.822***
(0.04)

0.102
(0.07)

0.103*
(0.06)

0.142**
(0.06)

0.147**
(0.06)

N
18

Y
18

0.197***
(0.01)
[0.039]
0.850***
(0.05)
[0.148]
0.157***
(0.03)
[0.129]
Y
5205

0.209***
(0.03)

0.808***
(0.05)

0.209***
(0.01)
[0.022]
0.874***
(0.05)
[0.149]
0.154***
(0.03)
[0.086]
N
5205

N
18

N
18

Appendix 3 Table 2: Usual weekly hours of work of domestic workers: Difference-in-differences, unweighted.
Excluding cusp wave

All waves
OLS

POST

Pre-law wage gap (WG)

Pre-law wage gap (WG)*POST

Controls for age, education, African?

N

Two-step estimator

OLS

(1)

(2)

(3)

(4)

(5)

(6)

(7)

Two-step estimator
(8)

0.512
(0.62)
[0.743]
5.208
(3.03)
[5.794]
2.695
(1.58)
[2.689]
N
6876

0.446
(0.59)
[0.709]
4.084
(2.46)
[4.452]
3.064*
(1.59)
[2.577]
Y
6876

0.557
(1.73)

0.471
(1.54)

0.516
(1.54)

5.219*
(2.44)

6.810**
(3.06)

5.910**
(2.66)

3.509
(4.10)

3.636
(3.77)

4.329
(4.31)

4.42
(3.91)

N
6876

Y
6876

0.664
(0.82)
[1.072]
4.426
(2.77)
[4.818]
3.536
(2.15)
[3.633]
Y
5824

0.622
(1.75)

5.990**
(2.76)

0.798
(0.88)
[1.151]
5.641
(3.52)
[6.234]
3.129
(2.25)
[3.931]
N
5824

N
18

Y
18

Robust standard errors presented in each column. In columns 1, 2, 5 and 6, standard errors are Eicker-White, clustered at the province level. Signicance at pb 0.001***, pb 0.05** or pb 0.01*
level. Critical values for signicance for two-step estimates are taken from the t-distribution (9-4=5 degrees of freedom). See text for details. Standard errors in square brackets are block bootstrapped, treating the province as the block. Sample in columns 14 include domestic workers in all waves; sample in columns 58 exclude domestic workers in the September 2002 cusp
wave.

T. Dinkelman, V. Ranchhod / Journal of Development Economics 99 (2012) 2745

References
Akerlof, George, 1982. Labor contracts as partial gift exchange. Quarterly Journal of
Economics 97 (4), 543569.
Akerlof, George, 1984. Gift exchange and efciency-wage theory: four views. American
Economic Review 74 (2), 7983.
Akerlof, George, Yellen, Janet, 1988. Fairness and unemployment. American Economic
Review 78 (2), 4449.
Ashenfelter, Orley, Smith, Richard, 1979. Compliance with the minimum wage law.
Journal of Political Economy 2, 333350 April.
Banerjee, Abhijit V., Duo, Esther, 2007. "The economic lives of the poor". Journal of
Economic Perspectives 21 (1), 141168.
Basu, Arnab K., Chau, Nancy H., Kanbur, Ravi, 2007. Turning a blind eye: costly enforcement, credible commitment and minimum wage laws. IZA Discussion Paper No.
2998. . August.
Bell, Linda A., 1997. The impact of minimum wages in Mexico and Colombia. Journal of
Labor Economics 15 (2).
Bertrand, Marianne, Duo, Esther, Mullainathan, Sendhil, 2004. How much should we
trust difference-in-differences estimates? Quarterly Journal of Economics 119 (1),
249275.
Besley, Timothy, Burgess, Robin, 2004. Can labor regulation hinder economic performance? Evidence from India. The Quarterly Journal of Economics 119 (1), 91134.
Bhorat, Haroon, Ravi, Kanbur, Natasha, Mayet, 2010. The determinants of minimum
wage violation in South Africa. Technical Report, Development Policy Research
Unit Working Paper. October.
Bloom, David, Grenier, Georges, 1986. Models of Firm Behavior under Minimum Wage
Legislation. NBER Working Paper. 1877.
Brandts, Jordi, Charness, Gary, 2004. Do labour market conditions affect gift exchange?
Some experimental evidence. Universitat Pompeu Fabra Working Paper No. 491.
Branson, Nicola, 2009. Re-weighting the OHS and LFS National Household Survey Data
to create a consistent series over time: A Cross Entropy Estimation Approach. SALDRU/DataFirst Working Paper Series number 38.
Brown, Charles, 1988. Minimum wage laws: are they overrated? Journal of Economic
Perspectives 2 (3).
Brown, Charles, 1999. Handbook of Labor Economics, Vol. 3. Elsevier Science.
Budlender, D., Kalula, E., Letsoela, B., Ntshalintshali, B., Rustomjee, Z., 2002. Recommendations of the Employment Conditions Commission on the Investigation into
the Domestic Worker Sector. Technical Report, South African Department of
Labour.
Card, David, Krueger, Alan, 1997. Myth and Measurement: The New Economics of the
Minimum Wage. Princeton University Press.
Chang, Yangming, 1992. Noncompliance behavior of risk averse rms under the minimum wage law. Public Finance Quarterly 20 (3), 390402.
Chang, Yangming, Ehrlich, Isaac, 1985. On the economics of compliance with the minimum wage law. Journal of Political Economy 93 (1), 8491.
Cock, Jacklyn, 1989. Maids and Madams: Domestic Workers Under Apartheid.
Women's Press, London.
Cortes, Kalena, 2004. Wage Effects on Immigrants from an Increase in the Minimum
Wage Rate: An Analysis by Immigrant Industry Concentration. Working Paper,
IZA Discussion Paper No. 1064.
Cortes, Patricia and Jessica Pan, Outsourcing household production: Foreign domestic
helpers and native labor supply in Hong Kong, 2009. Mimeo, Chicago Booth
School.
DiNardo, John, Fortin, Nicole, Lemieux, Thomas, 1996. Labor market institutions and
the distribution of wages, 19731992: a semiparametric analysis. Econometrica
64, 10011044 September.
Donald, Stephen G., Lang, Kevin, 2007. Inference with difference-in-differences and
other panel data. The Review of Economics and Statistics 89 (2), 221233.
Falk, Armin, Fehr, Ernst, Fischbacher, Urs, 2000. Testing Theories of Fairness-Intentions
Matter. University of Zurich.

45

Falk, Armin, Fehr, Ernst, Zehnder, Christian, 2005. The behavioral effects of the minimum wage. IZA Discussion Paper No. 1625. . June.
Fehr, Ernst, Simon, Gachter, Georg, Kirschteiger, 1997. Reciprocity as a contract enforcement device: experimental evidence. Econometrica 65 (4).
Gindling, T.H., Katherine, Terrell, 2004. Legal minimum wages and the wages of formal
and informal sector workers in Costa Rica. IZA Discussion Paper No. 1018. .
February.
Grenier, Georges, 1982. On compliance with the minimum wage law. Journal of Political Economy 90 (1), 184187.
Hertz, Tom, 2005. The effect of minimum wages on the employment and earnings of
South Africa's domestic service workers. University of Cape Town, Development
Policy Research Unit Working Paper 05/99.
Konow, James, 2003. Which is the fairest one of all? A positive analysis of justice theories. Journal of Economic Literature 41 (4).
Kremer, Michael and Stanley Watt, The globalization of household production, September 2006. Mimeo, Harvard University.
Lee, David, 1999. Wage inequality in the United States during the 1980s: rising dispersion or falling minimum wage? Quarterly Journal of Economics 114 (3), 9771023
August.
Lemos, Sara, 2009. Minimum wage effects in a developing country. Labour Economics
16.
Lott, John R., Roberts, Russell D., 1995. The expected penalty for committing a crime: an
analysis of minimum wage violations. Journal of Human Resources 30 (2),
397408.
Louw, Megan, van der Berg, Servaas, 2004. Labour trends observed in South Africa:
19952002. Technical Report, The World Bank and University of Stellenbosch.
DIFID-WB Collaboration on Knowledge and Skills in the New Economy.
Lustig, Nora, McLeod, Daniel, 1997. Labor Markets in Latin America. Brookings Institution Press.
Maloney, W., Menendez, J., 2004. Law and Employment: Lessons from Latin America
and the Caribbean. NBER and University of Chicago.
Mas, Alex, 2006. Pay, reference points and policy performance. The Quarterly Journal of
Economics 121 (3), 783821.
Neumark, David, Wascher, William, 1995. Minimum wage effects on school and work
transitions of teenagers. American Economic Review Papers and Proceedings 85
(2), 244249 May.
Neumark, David, Wascher, William, 2007. Minimum wages and employment. IZA Discussion Paper No. 2570. . January.
Rebitzer, James B., Taylor, Lowell J., 1995. The consequences of minimum wage laws:
some new theoretical ideas. Journal of Public Economics 56, 245255.
Rubinow, I.M., 1906. The problem of domestic service. Journal of Political Economy 14
(8), 502519.
South African Department of Labor, 1997. Basic Conditions of Employment Act 1997.
South African Government.
Squire, Lyn, Suthiwart-Narueput, Sethaput, 1997. The impact of labor market regulations. World Bank Research Observer 11 (1), 119143.
Statistics South Africa, 20002003. Labor Force Surveys Waves 48. Technical Report,
South African Government.
Thies, Clifford F., 1991. The rst minimum wage laws. The Cato Journal 10 (3),
715746.
Weil, David, 2005. Public enforcement/private monitoring: evaluating a new approach
to regulating the minimum wage. Industrial & Labor Relations Review 58 January.
Yamada, Hiroyuki, Four essays on labor and development economics: Chapter 1, The
impact of the introduction of sectoral minimum wages on low wage markets in a
low income country: evidence from South Africa. PhD dissertation, University of
Chicago August 2008.